Offer the full playback rate list on Safari (#24444)

* Rate events over at least one second

EventsPerSecond.eps() divided the event count by the time since start(),
which can be a few milliseconds right after a restart. Frames buffered
during an ffmpeg restart then report as 100+ fps, and the same happens to
the detector fps. Use a window of at least one second.

* Keep sub-second windows consistent

Floor the divisor at the window length when the window is shorter than a
second, so a caller with a sub-second window still gets its true rate.

CODEOWNERS

@@ -5,3 +5,4 @@
 /docker/rockchip/ @MarcA711
 /docker/rocm/ @harakas
 /docker/hailo8l/ @spanner3003
+/docker/deepx/ @sixfab

docker/deepx/user_installation.sh

@@ -0,0 +1,131 @@
+#!/bin/bash
+
+# Installs the DEEPX NPU kernel driver and the DX-RT runtime on the Docker host,
+# then enables the vendor's dxrt.service. A container cannot load kernel
+# modules, so this runs outside the image; the driver creates the /dev/dxrt*
+# nodes and the daemon multiplexes the NPU across host and container.
+#
+# Driver, runtime and firmware versions must agree or inference hangs instead
+# of failing at startup. The set this script installs is pinned in
+# driver_version, runtime_version and firmware_version below; move them
+# together, never one at a time.
+#
+# DEEPX NPU support in Frigate is maintained by Sixfab (https://sixfab.com).
+
+set -euo pipefail
+
+driver_version="v2.6.0"
+# the commit the tag resolves to, since DEEPX signs neither tags nor releases
+# and this is compiled and installed as root. Update both together
+driver_commit="7074748e7104f470b02f517583abba652b3f05fa"
+firmware_version="v2.7.4"
+
+sudo apt-get update
+sudo apt-get install -y git build-essential "linux-headers-$(uname -r)" pciutils wget
+
+if ! lspci -d 1ff4: | grep -q .; then
+    echo "No DEEPX device found on the PCIe bus (lspci -d 1ff4:)."
+    echo "Check that the module is seated correctly before continuing."
+    exit 1
+fi
+
+# fetch the pinned commit rather than cloning the tag, so a retag cannot swap
+# in different source. The build directory is reused so a second run after a
+# failure does not stop on the directory already being there
+mkdir -p dx_rt_npu_linux_driver
+cd dx_rt_npu_linux_driver
+git init -q
+git remote get-url origin > /dev/null 2>&1 ||
+    git remote add origin https://github.com/DEEPX-AI/dx_rt_npu_linux_driver.git
+git fetch --depth 1 origin "${driver_commit}"
+git checkout -q FETCH_HEAD
+
+fetched_commit=$(git rev-parse HEAD)
+if [[ "${fetched_commit}" != "${driver_commit}" ]]; then
+    echo "Fetched commit ${fetched_commit} does not match pinned driver_commit ${driver_commit}."
+    echo "Refusing to build unverified driver source."
+    exit 1
+fi
+
+cd modules
+
+sudo ./build.sh -c install --reload
+
+sudo depmod -A
+
+# dx_dma is the PCIe transport, dxrt_driver the NPU driver on top of it
+for module in dx_dma dxrt_driver; do
+    if ! sudo modprobe "${module}"; then
+        echo "Unable to load the ${module} kernel module, common reasons are:"
+        echo "- Secure Boot is enabled and is rejecting the unsigned module."
+        echo "- The running kernel does not match the installed linux-headers."
+        exit 1
+    fi
+done
+
+if ! compgen -G "/dev/dxrt*" > /dev/null; then
+    echo "Modules loaded but no /dev/dxrt* device node appeared."
+    echo "Run ./sanity_check.sh from the driver repo to diagnose."
+    exit 1
+fi
+
+runtime_version="v3.4.0"
+declare -A runtime_sha256=(
+    [amd64]="736cfef009ce9e974ab1ab610d867239d19d72a426a53e367ddcbd53297b6e20"
+    [arm64]="eb6107f5f02f2ad76ae89f414e8b5f346f34fbc6f0888236136853a26be6f6a0"
+)
+
+runtime_release="${runtime_version#v}"
+deb_arch=$(dpkg --print-architecture)
+deb_file="/tmp/libdxrt-bin_${runtime_release}_${deb_arch}.deb"
+
+wget -qO "${deb_file}" \
+    "https://raw.githubusercontent.com/DEEPX-AI/dx_rt/${runtime_version}/release/${runtime_release}/libdxrt-bin_${runtime_release}_${deb_arch}.deb"
+
+expected_sha256="${runtime_sha256[${deb_arch}]:-}"
+if [[ -z "${expected_sha256}" ]]; then
+    echo "No pinned SHA-256 for architecture ${deb_arch}; refusing to install."
+    exit 1
+fi
+if [[ "$(sha256sum "${deb_file}" | cut -d' ' -f1)" != "${expected_sha256}" ]]; then
+    echo "SHA-256 mismatch for ${deb_file}; refusing to install."
+    exit 1
+fi
+
+sudo dpkg -i "${deb_file}"
+sudo ldconfig
+rm -f "${deb_file}"
+
+sudo cp /usr/share/libdxrt-bin/service/dxrt.service /etc/systemd/system/
+
+# With an endpoint set, dxrtd binds that path only, so the socket goes in a
+# directory Frigate can mount (kept across restarts so the mount stays valid)
+# and a symlink at the default /tmp path keeps host tools that do not set the
+# variable working through their own fallback.
+sudo mkdir -p /etc/systemd/system/dxrt.service.d
+sudo tee /etc/systemd/system/dxrt.service.d/frigate.conf > /dev/null <<'UNIT'
+[Service]
+RuntimeDirectory=dxrt
+RuntimeDirectoryMode=0755
+RuntimeDirectoryPreserve=yes
+Environment=DXRT_DYNAMIC_IPC_ENDPOINT=/run/dxrt/dxrt_dynamic_ipc.sock
+ExecStartPost=/bin/ln -sfn /run/dxrt/dxrt_dynamic_ipc.sock /tmp/dxrt_dynamic_ipc.sock
+UNIT
+
+sudo systemctl daemon-reload
+sudo systemctl enable dxrt.service
+sudo systemctl restart dxrt.service
+
+if ! sudo systemctl is-active --quiet dxrt.service; then
+    echo "dxrt.service did not start. Check: sudo journalctl -u dxrt.service"
+    exit 1
+fi
+
+echo "DEEPX driver and runtime installation complete."
+echo "Driver version:   $(modinfo -F version dxrt_driver) (expected ${driver_version#v})"
+echo "Runtime version:  ${runtime_release}"
+echo "Device node(s):   $(echo /dev/dxrt*)"
+echo
+echo "This driver expects NPU firmware ${firmware_version}. Check it with:"
+echo "  dxrt-cli --status"
+echo "Update the module if it does not match before starting Frigate."

docker/main/rootfs/etc/s6-overlay/s6-rc.d/init-devices/run

@@ -37,6 +37,7 @@ device_globs=(
     "/dev/nvmap"
     "/dev/nvidia*"
     "/dev/memx*"
+    "/dev/dxrt*"
 )
 
 IFS=',' read -ra extra_globs <<< "${DEVICE_ACL_PATHS:-}"

docs/data/object_detectors_models.yaml

@@ -967,6 +967,65 @@ memryx:
             # The .zip file must contain:
             # ├── ssdlite_mobilenet.dfp          (a file ending with .dfp)
             # └── ssdlite_mobilenet_post.onnx    (optional; only if the model includes a cropped post-processing network)
+deepx:
+  title: DEEPX NPU
+  models:
+    - key: yolo
+      label: YOLO
+      recommended: true
+      download: No model is bundled with Frigate. Download a pre-compiled YOLO `.dxnn` model from the [DEEPX ModelZoo](https://developer.deepx.ai/modelzoo) or compile your own with DX-COM, then bind-mount it into the container and point the model's `path` at it. The recommended model is `yolox-s_640x640_ppu.dxnn`. Its Post-Processing Unit (PPU) compile moves candidate selection onto the NPU, which makes it the fastest ModelZoo model measured through Frigate (about 13 ms on a DX-M1). The output layout is read from the compiled model, so anchor-based, anchor-free, NMS-in-head and PPU models (anchor-based or anchor-free) all need no extra configuration; prefer a `PPU` variant whenever the ModelZoo offers one. PPU models must be compiled with DX-COM 2.4.0 or later, which writes the head layout Frigate reads into the file.
+      ui: |-
+        Navigate to **Settings > System > Detection models** and select **DEEPX NPU** from the **Hardware** dropdown. Then, on the same model, open the **Custom Model** tab and configure:
+
+        | Field                                    | Value                                                |
+        | ---------------------------------------- | ---------------------------------------------------- |
+        | **Custom object detector model path**    | `/config/model_cache/deepx/yolox-s_640x640_ppu.dxnn` |
+        | **Label map for custom object detector** | `/labelmap/coco-80.txt`                              |
+        | **Object detection model input width**   | `640`                                                |
+        | **Object detection model input height**  | `640`                                                |
+        | **Model Input Pixel Color Format**       | `rgb` (Frigate's default value)                      |
+        | **Model Input Tensor Shape**             | `nhwc` (Frigate's default value)                     |
+        | **Model Input D Type**                   | `int` (Frigate's default value)                      |
+        | **Object Detection Model Type**          | `yolo-generic`                                       |
+
+        Quantization is baked into the compiled model, so no normalization is applied on the host and the input defaults do not need to be overridden.
+      yaml: |-
+        models:
+          - devices:
+              - deepx:PCIe:0
+            path: /config/model_cache/deepx/yolox-s_640x640_ppu.dxnn
+            labelmap_path: /labelmap/coco-80.txt
+            model_type: yolo-generic
+            width: 640
+            height: 640
+    - key: yolox
+      label: YOLOX
+      recommended: false
+      download: No model is bundled with Frigate. Download a pre-compiled YOLOX `.dxnn` model from the [DEEPX ModelZoo](https://developer.deepx.ai/modelzoo), then bind-mount it into the container and point the model's `path` at it. The `_ppu` variant is faster and also works with the `yolo-generic` model type; the plain export needs `yolox` so its raw head is decoded.
+      ui: |-
+        Navigate to **Settings > System > Detection models** and select **DEEPX NPU** from the **Hardware** dropdown. Then, on the same model, open the **Custom Model** tab and configure:
+
+        | Field                                    | Value                                    |
+        | ---------------------------------------- | ---------------------------------------- |
+        | **Custom object detector model path**    | `/config/model_cache/deepx/yolox-s_640x640.dxnn`|
+        | **Label map for custom object detector** | `/labelmap/coco-80.txt`                  |
+        | **Object detection model input width**   | `640`                                    |
+        | **Object detection model input height**  | `640`                                    |
+        | **Model Input Pixel Color Format**       | `rgb` (Frigate's default value)          |
+        | **Model Input Tensor Shape**             | `nhwc` (Frigate's default value)         |
+        | **Model Input D Type**                   | `int` (Frigate's default value)          |
+        | **Object Detection Model Type**          | `yolox`                                  |
+
+        The width and height must match the resolution the `.dxnn` file was compiled for.
+      yaml: |-
+        models:
+          - devices:
+              - deepx:PCIe:0
+            path: /config/model_cache/deepx/yolox-s_640x640.dxnn
+            labelmap_path: /labelmap/coco-80.txt
+            model_type: yolox
+            width: 640
+            height: 640
 tensorrt:
   title: TensorRT
   models:

docs/docs/configuration/advanced/reference.md

@@ -800,7 +800,7 @@ lpr:
 # to Google or OpenAI's LLMs to generate descriptions. GenAI features can be configured at
 # the camera level to enhance privacy for indoor cameras.
 # NOTE: genai is a map of named providers. Each key is a name you choose for the provider,
-#       and each role (chat, descriptions, embeddings) may be assigned to exactly one provider.
+#       and each role (chat, descriptions, embeddings, transcribe) may be assigned to exactly one provider.
 genai:
   # Required: name of the provider (chosen by you, used to reference it elsewhere)
   my_provider:
@@ -813,11 +813,13 @@ genai:
     # Required: The model to use with the provider.
     model: gemini-1.5-flash
     # Optional: Roles this provider handles (default: shown below)
-    # Each role (chat, descriptions, embeddings) must be assigned to exactly one provider.
+    # Each role (chat, descriptions, embeddings, transcribe) must be assigned to exactly
+    # one provider.
     roles:
       - chat
       - descriptions
       - embeddings
+      - transcribe
     # Optional additional args to pass to the GenAI Provider (default: None)
     provider_options:
       keep_alive: -1
@@ -830,13 +832,19 @@ genai:
 audio_transcription:
   # Optional: Enable live and speech event audio transcription (default: shown below)
   enabled: False
+  # Optional: The transcription backend (default: shown below)
+  # Either 'whisper' for Frigate's built-in local models, or the name of a genai
+  # provider that has 'transcribe' in its roles. device and model_size are ignored
+  # when a genai provider is named.
+  model: whisper
   # Optional: The device to run the models on for live transcription. (default: shown below)
   device: CPU
   # Optional: Set the model size used for live transcription. (default: shown below)
   model_size: small
   # Optional: Set the language used for transcription translation. (default: shown below)
-  # List of language codes: https://github.com/openai/whisper/blob/main/whisper/tokenizer.py#L10
-  language: en
+  # Use 'auto' to let the model detect the language, or a language code from
+  # https://github.com/openai/whisper/blob/main/whisper/tokenizer.py#L10
+  language: auto
 
 # Optional: Configuration for classification models
 classification:

docs/docs/configuration/audio_detectors.md

@@ -204,7 +204,7 @@ Frequently-heard labels like `speech` can generate a lot of events, and each eve
 
 ### Audio Transcription
 
-Frigate supports fully local audio transcription using either `sherpa-onnx` or OpenAI's open-source Whisper models via `faster-whisper`. The goal of this feature is to support Semantic Search for `speech` audio events. Frigate is not intended to act as a continuous, fully-automatic speech transcription service. Automatically transcribing all speech (or queuing many audio events for transcription) requires substantial CPU (or GPU) resources and is impractical on most systems. For this reason, transcriptions for events are initiated manually from the UI or the API rather than being run continuously in the background.
+Frigate supports fully local audio transcription using either `sherpa-onnx` or OpenAI's open-source Whisper models via `faster-whisper`, and can alternatively offload transcription to a [GenAI provider](#genai-provider). The goal of this feature is to support Semantic Search for `speech` audio events. Frigate is not intended to act as a continuous, fully-automatic speech transcription service. Automatically transcribing all speech (or queuing many audio events for transcription) requires substantial CPU (or GPU) resources and is impractical on most systems. For this reason, transcriptions for events are initiated manually from the UI or the API rather than being run continuously in the background.
 
 :::info
 
@@ -224,6 +224,7 @@ To enable transcription, configure it globally and optionally disable for specif
 **Global:** Navigate to <NavPath path="Settings > Enrichments > Audio transcription" />.
 
 - Set **Enable audio transcription** to on
+- Set **Audio transcription model or GenAI provider name** to `whisper` for Frigate's built-in local models, or to the name of a GenAI provider
 - Set **Transcription device** to the desired device
 - Set **Model size** to the desired size
 
@@ -235,6 +236,7 @@ To enable transcription, configure it globally and optionally disable for specif
 ```yaml
 audio_transcription:
   enabled: True
+  model: whisper
   device: ...
   model_size: ...
 ```
@@ -263,20 +265,88 @@ The optional config parameters that can be set at the global level include:
 - **`enabled`**: Enable or disable the audio transcription feature.
   - Default: `False`
   - It is recommended to only configure the features at the global level, and enable it at the individual camera level.
+- **`model`**: The transcription backend.
+  - Default: `whisper`
+  - `whisper` uses Frigate's built-in local models, described by `device` and `model_size` below.
+  - Any other value must name a key in your `genai` config whose entry has `transcribe` in its `roles`. See [GenAI Provider](#genai-provider).
 - **`device`**: Device to use to run transcription and translation models.
   - Default: `CPU`
   - This can be `CPU` or `GPU`. The `sherpa-onnx` models are lightweight and run on the CPU only. The `whisper` models can run on GPU but are only supported on CUDA hardware.
+  - Ignored when `model` names a GenAI provider.
 - **`model_size`**: The size of the model used for live transcription.
   - Default: `small`
   - This can be `small` or `large`. The `small` setting uses `sherpa-onnx` models that are fast, lightweight, and always run on the CPU but are not as accurate as the `whisper` model.
-  - This config option applies to **live transcription only**. Recorded `speech` events will always use a different `whisper` model (and can be accelerated for CUDA hardware if available with `device: GPU`).
-- **`language`**: Defines the language used by `whisper` to translate `speech` audio events (and live audio only if using the `large` model).
-  - Default: `en`
-  - You must use a valid [language code](https://github.com/openai/whisper/blob/main/whisper/tokenizer.py#L10).
+  - This config option applies to **live transcription only**. With `model: whisper`, recorded `speech` events always use a different `whisper` model (and can be accelerated for CUDA hardware if available with `device: GPU`).
+  - Ignored when `model` names a GenAI provider.
+- **`language`**: Defines the language used to transcribe and translate `speech` audio events (and live audio only if using the `large` model or a GenAI provider).
+  - Default: `auto`
+  - `auto` lets the model detect the language itself, which most models do well. Set an explicit language only if detection is picking the wrong one.
+  - Otherwise you must use a valid [language code](https://github.com/openai/whisper/blob/main/whisper/tokenizer.py#L10).
   - Transcriptions for `speech` events are translated.
   - Live audio is translated only if you are using the `large` model. The `small` `sherpa-onnx` model is English-only.
 
-The only field that is valid at the camera level is `enabled`.
+The only field that is valid at the camera level is `enabled`. In particular `model` is global only: the transcription backend is a process-wide resource shared by every camera.
+
+#### GenAI Provider
+
+Frigate can send audio to a GenAI provider for transcription when that provider has the `transcribe` role. This is useful if you already run a GenAI provider, or if you do not have the CPU/GPU headroom for a local whisper model. Supported providers are **OpenAI**, **Azure OpenAI**, **Gemini**, and **llama.cpp** with an audio-capable model (a dedicated ASR model such as Qwen3-ASR, or a general multimodal model that accepts audio). Ollama is not supported as it has no audio input.
+
+To use a GenAI provider for audio transcription:
+
+1. Configure a GenAI provider with `transcribe` in its `roles`.
+2. Set the audio transcription model to that GenAI config key (e.g. `whisper_cloud`).
+
+<ConfigTabs>
+<TabItem value="ui">
+
+Navigate to <NavPath path="Settings > Enrichments > Audio transcription" />.
+
+| Field                                                | Description                                                                                              |
+| ---------------------------------------------------- | ---------------------------------------------------------------------------------------------------------- |
+| **Audio transcription model or GenAI provider name** | Set to the GenAI config key (e.g. `whisper_cloud`) to use a configured GenAI provider for transcription |
+
+The GenAI provider must also be configured with the `transcribe` role under <NavPath path="Settings > Enrichments > Generative AI" />.
+
+</TabItem>
+<TabItem value="yaml">
+
+```yaml
+genai:
+  whisper_cloud:
+    provider: openai
+    api_key: your-api-key
+    model: gpt-transcribe
+    roles:
+      - transcribe
+
+audio_transcription:
+  enabled: True
+  model: whisper_cloud
+  language: en
+```
+
+</TabItem>
+</ConfigTabs>
+
+:::warning
+
+**Give `transcribe` its own `genai` entry.** A `genai` entry has a single `model` string that is shared by every role it holds, so `roles: [descriptions, transcribe]` would send the same model name to both the chat endpoint and the transcription endpoint. Transcription models and chat models are almost never the same model, so define a dedicated entry as shown above.
+
+:::
+
+:::warning
+
+**Live transcription against a metered provider is billed continuously.** In live mode Frigate uploads an overlapping ~2 second window of audio roughly once per second, per camera, for as long as audio stays above that camera's `audio.min_volume`. Windows below that threshold are never uploaded, which is what keeps a quiet camera near zero requests, but a camera pointed at a busy street will keep sending.
+
+Three things keep this opt-in: `transcribe` is not one of the default roles, live transcription is off by default, and the volume gate suppresses silence. Transcription of recorded `speech` events is unaffected - it remains a manual, one-request-per-event action.
+
+:::
+
+`device` and `model_size` have no effect on this path and no local model is ever downloaded.
+
+`language` defaults to `auto`, which sends no language hint and lets the model detect it. Most audio models detect language well, so leave it on `auto` unless detection is picking the wrong one.
+
+When set explicitly, it is sent as the transcription endpoint's native `language` parameter for OpenAI, Azure, and llama.cpp, and as part of the prompt for Gemini. This matters for dedicated ASR models such as Qwen3-ASR: they read the prompt as contextual biasing rather than as an instruction, so a language named in the prompt is ignored, while the endpoint parameter is honored.
 
 #### Live transcription
 
@@ -292,6 +362,8 @@ Results can be error-prone due to a number of factors, including:
 
 For speech sources close to the camera with minimal background noise, use the `small` model.
 
+A [GenAI provider](#genai-provider) is generally the most accurate option for live transcription, at the cost of a network round trip per window. That round trip has to stay under about a second to keep up with the audio; if it does not, Frigate drops the oldest buffered audio rather than letting the backlog grow.
+
 If you have CUDA hardware, you can experiment with the `large` `whisper` model on GPU. Performance is not quite as fast as the `sherpa-onnx` `small` model, but live transcription is far more accurate. Using the `large` model with CPU will likely be too slow for real-time transcription.
 
 #### Transcription and translation of `speech` audio events
@@ -308,7 +380,7 @@ Only one `speech` event may be transcribed at a time. Frigate does not automatic
 
 :::
 
-Recorded `speech` events will always use a `whisper` model, regardless of the `model_size` config setting. Without a supported Nvidia GPU, generating transcriptions for longer `speech` events may take a fair amount of time, so be patient.
+With `model: whisper`, recorded `speech` events always use a `whisper` model, regardless of the `model_size` config setting. Without a supported Nvidia GPU, generating transcriptions for longer `speech` events may take a fair amount of time, so be patient. With a [GenAI provider](#genai-provider), the recorded clip is sent to the provider instead and no local model is used.
 
 #### FAQ
 

docs/docs/configuration/genai/config.md

@@ -43,7 +43,7 @@ genai:
 
 The examples on this page all use `my_provider`, but the name is arbitrary and is only used to reference the provider elsewhere in the config (for example, `semantic_search.model`).
 
-Each provider handles one or more **roles**: `chat`, `descriptions`, and `embeddings`. A provider handles all three by default, and each role may be assigned to exactly one provider. Define a single provider if you want it to do everything, or split the roles across several providers using the `roles` option.
+Each provider handles one or more **roles**: `chat`, `descriptions`, `embeddings`, and `transcribe`. A provider handles the first three by default; `transcribe` must always be listed explicitly, and is not available on Ollama, which has no audio input. Each role may be assigned to exactly one provider. Define a single provider if you want it to do everything, or split the roles across several providers using the `roles` option.
 
 If the provider you choose requires an API key, you may either directly paste it in your configuration, or store it in an environment variable prefixed with `FRIGATE_`.
 
@@ -63,11 +63,11 @@ Running Generative AI models on CPU is not recommended, as high inference times
 
 You must use a vision-capable model with Frigate. The following models are recommended for local deployment of the `descriptions` and `chat` roles:
 
-| Model               | Notes                                                                                                                                                                |
-| ------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `qwen3-vl`          | Strong visual and situational understanding, enhanced ability to identify smaller objects and interactions with object.                                              |
-| `qwen3.6`/`qwen3.8` | Strong situational understanding, but missing DeepStack from qwen3-vl leading to worse performance for identifying objects in people's hand and other small details. |
-| `gemma4`            | Strong situational understanding, sometimes resorts to more vague terms like 'interacts' instead of assigning a specific action.                                     |
+| Model               | Review [frame mode](/configuration/genai/genai_review#frame-mode) | Notes                                                                                                                                                                |
+| ------------------- | --------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `qwen3-vl`          | `frames`                                                              | Strong visual and situational understanding, enhanced ability to identify smaller objects and interactions with object. Follows a sequence of frames on its own.     |
+| `qwen3.6`/`qwen3.8` | `frames`                                                              | Strong situational understanding, but missing DeepStack from qwen3-vl leading to worse performance for identifying objects in people's hand and other small details. |
+| `gemma4`            | `annotated_frames`                                                    | Strong situational understanding, sometimes resorts to more vague terms like 'interacts' instead of assigning a specific action. Loses track of activity that repeats or reverses, so it benefits from annotated frames. |
 
 #### Embedding models
 
@@ -77,6 +77,17 @@ The `embeddings` role needs a different kind of model. Text queries are matched
 | -------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | `qwen3-vl-embedding` | Multimodal embeddings for [Semantic Search](/configuration/semantic_search#genai-provider). Must be served by llama.cpp started with `--embeddings` and `--mmproj`. |
 
+#### Transcription models
+
+The `transcribe` role needs a model that accepts audio input. A text-only or vision-only model cannot serve this role. The following are recommended for local deployment of the `transcribe` role:
+
+| Model       | Notes                                                                                                                                                                                                      |
+| ----------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| `qwen3-asr` | Dedicated speech recognition model covering 30 languages, and the better choice for transcription quality. It only transcribes, so it cannot be shared with the `descriptions` or `chat` roles.            |
+| `gemma4`    | General multimodal model that accepts audio as well as images, so one served model can cover `transcribe` alongside the other roles. Transcript quality is below `qwen3-asr`, particularly on noisy audio. |
+
+Both must be served by llama.cpp started with the matching audio `--mmproj`. llama.cpp only reports audio support when an audio projector is loaded. Without it Frigate sees the model as text-only and the `transcribe` role is unavailable in the UI. Frigate transcribes through the server's `/v1/audio/transcriptions` route, which llama.cpp serves for any audio-capable model.
+
 :::info
 
 Each model is available in multiple parameter sizes (3b, 4b, 8b, etc.). Larger sizes are more capable of complex tasks and understanding of situations, but requires more memory and computational resources. It is recommended to try multiple models and experiment to see which performs best.

docs/docs/configuration/genai/review_summaries.md

@@ -192,6 +192,43 @@ review:
 </TabItem>
 </ConfigTabs>
 
+### Frame Mode
+
+Review items are sent to the model as a sequence of still frames. Some models follow that sequence well on their own; others lose track of activity that repeats or reverses, and describe a single trip when the subject actually made several. The `frame_mode` option controls how those frames are presented.
+
+- `frames` (default): the prompt followed by the frames, exactly as earlier versions of Frigate sent them.
+- `annotated_frames`: each frame is preceded by its frame number and elapsed time, along with notes describing what the object tracker recorded at that moment, such as an object being first detected, starting to move, turning around, stopping, or no longer being detected.
+
+The notes come from tracking data rather than from the images, so they describe activity the model may not have picked up on its own. In testing with a person carrying three waste bins to the curb one at a time, `gemma4` described a single trip on every attempt with `frames`, and consistently described multiple trips with `annotated_frames`. Models that already handle these sequences well, such as the `qwen3-vl` family, gain little and should stay on `frames`.
+
+Annotated mode also caps the number of frames, since the notes already establish the order of events and extra near-duplicate frames tend to crowd out the middle of a clip. Longer review items are sampled more sparsely as a result, and typically use fewer tokens than `frames` mode for the same item.
+
+:::note
+
+Annotated mode needs tracking data for the review item. If none is available, Frigate falls back to sending plain frames for that item.
+
+:::
+
+<ConfigTabs>
+<TabItem value="ui">
+
+Navigate to <NavPath path="Settings > Global configuration > Review" />.
+
+- Set **GenAI config > Frame mode** to the desired mode (e.g., `annotated_frames`)
+
+</TabItem>
+<TabItem value="yaml">
+
+```yaml {4}
+review:
+  genai:
+    enabled: true
+    frame_mode: annotated_frames
+```
+
+</TabItem>
+</ConfigTabs>
+
 ### Response Style
 
 Different models respond to the built-in prompt with very different writing styles: some produce natural narration while others sound short and mechanical. The `response_style` option selects a writing style preset that rewords the prompt's instructions for the user-facing fields (the title, short summary, and scene description). Presets replace those instructions rather than adding extra ones, so the model never receives competing style directions.

docs/docs/configuration/object_detectors.md

@@ -24,6 +24,7 @@ Frigate supports multiple different detectors that work on different types of ha
 - [Coral EdgeTPU](#edge-tpu-detector): The Google Coral EdgeTPU is available in USB, Mini PCIe, and m.2 formats allowing for a wide range of compatibility with devices.
 - [Hailo](#hailo): The Hailo-8, Hailo-8L and Hailo-8R AI Acceleration modules are available in m.2 format with a HAT for RPi devices, offering a wide range of compatibility with devices.
 - <CommunityBadge /> [MemryX](#memryx-mx3): The MX3 Acceleration module is available in m.2 format, offering broad compatibility across various platforms.
+- <CommunityBadge /> [DEEPX](#deepx-npu): The DEEPX NPU is available in m.2 format and as a HAT+ for the Raspberry Pi 5, offering broad compatibility across various platforms.
 
 **AMD**
 
@@ -612,6 +613,63 @@ For detailed instructions on compiling models, refer to the [MemryX Compiler](ht
 
 ---
 
+## DEEPX NPU
+
+This detector is available for use with the DEEPX NPU, both the DX-M1 M.2 module and the DX-M1M on the Sixfab AI HAT+ for the Raspberry Pi 5. The configuration below applies unchanged to either form factor. DEEPX NPU support in Frigate is developed and maintained by [Sixfab](https://sixfab.com).
+
+See the [installation docs](../frigate/installation.md#deepx-npu) for information on installing the DEEPX kernel driver and runtime on the host and passing the NPU through to the container.
+
+To run a model on a DEEPX NPU, list a `deepx` device on that model.
+
+:::info
+
+The DX-RT Python bindings are not part of the Frigate image. They are downloaded and installed into `/config/.local` the first time a DEEPX device is configured, verified against pinned checksums, and updated automatically when a Frigate release pins a new version. If the container has no internet access, see [Detector runtimes](/frigate/network_requirements#detector-runtimes) for how to provide the files yourself.
+
+:::
+
+### Configuration {#configuration-deepx}
+
+<ModelConfigDropdown detectorTitle="DEEPX" models={objectDetectorsModels.deepx.models} />
+
+Frigate does not bundle a model for this detector. Models must be compiled to DEEPX's `.dxnn` format. Two model types are supported:
+
+- `yolo-generic` for YOLO object detection models, the recommended default. The detector reads the model's output layout from the compiled file, so anchor-based, anchor-free and NMS-in-head models all work with the same configuration, as do models compiled with DEEPX's Post-Processing Unit (PPU) support.
+- `yolox` for YOLOX models compiled without PPU support, whose raw head needs Frigate's YOLOX decoder. A YOLOX model compiled with PPU support works under either `yolox` or `yolo-generic`.
+
+The quickest way to get one is the [DEEPX ModelZoo](https://developer.deepx.ai/modelzoo), which publishes pre-compiled `.dxnn` files for a range of YOLO object detection models. Download the `.dxnn`, bind-mount it into the container, and point the model's `path` at it. Alternatively, compile your own model with the DX-COM compiler. The recommended starting point is `yolox-s_640x640_ppu.dxnn`, the fastest ModelZoo model measured through Frigate:
+
+```yaml
+models:
+  - devices:
+      - deepx:PCIe:0
+    path: /config/model_cache/deepx/yolox-s_640x640_ppu.dxnn
+    labelmap_path: /labelmap/coco-80.txt
+    model_type: yolo-generic
+    width: 640
+    height: 640
+```
+
+For PPU models, use a `.dxnn` compiled with DX-COM 2.4.0 or later. Frigate reads the PPU head layout the compiler writes into the file and refuses to load a PPU model without it.
+
+`model_type` must be set to `yolo-generic` or `yolox` to match the model; `yolo-generic` is the recommended default unless the model is a raw YOLOX export. Frigate defaults it to `ssd`, which this detector does not support, so the detector refuses to start on a model that leaves it unset.
+
+`width` and `height` must match the resolution the model was compiled for. Quantization parameters are baked into the `.dxnn` file at compile time, so no normalization is applied on the host and Frigate's default `input_tensor`, `input_pixel_format`, and `input_dtype` values do not need to be overridden.
+
+A DEEPX device is `PCIe:<index>`, as reported on the detector settings page. The NPU daemon multiplexes across processes, so the same device may be listed more than once to run additional inference processes against it:
+
+```yaml
+models:
+  - devices:
+      - deepx:PCIe:0
+      - deepx:PCIe:0
+```
+
+#### Label maps
+
+The object detection models in the DEEPX ModelZoo are trained on the standard 80-class COCO label set, so `labelmap_path` must be set to `/labelmap/coco-80.txt`. Frigate's default label map uses an extended 91-class COCO scheme, and leaving it in place will cause detections to be reported as the wrong object type. For `yolo-generic` models the label map is also what the detector uses to tell the output layout, so a label map with the wrong number of classes is reported as an error at startup.
+
+---
+
 ## NVidia TensorRT Detector
 
 Nvidia Jetson devices may be used for object detection using the TensorRT libraries. Due to the size of the additional libraries, this detector is only provided in images with the `-tensorrt-jp6` tag suffix, e.g. `ghcr.io/blakeblackshear/frigate:stable-tensorrt-jp6`. This detector is designed to work with Yolo models for object detection.

docs/docs/frigate/hardware.md

@@ -65,6 +65,11 @@ Frigate supports multiple different detectors that work on different types of ha
   - [Supports many model architectures](../../configuration/object_detectors#memryx-mx3)
   - Runs best with tiny, small, or medium-size models
 
+- <CommunityBadge /> [DEEPX](#deepx-npu): The DEEPX NPU is available in m.2 format and as a HAT+ for the Raspberry Pi 5, allowing for a wide range of compatibility with devices.
+  - [Supports YOLO model architectures](../../configuration/object_detectors#deepx-npu)
+  - Runs best with tiny or small size models
+  - Runs efficiently on low power hardware
+
 **AMD**
 
 - [ROCm](#rocm---amd-gpu): ROCm can run on AMD Discrete GPUs to provide efficient object detection
@@ -257,6 +262,32 @@ The MX3 is a pipelined architecture, where the maximum frames per second support
 
 Inference speeds may vary depending on the host platform. The above data was measured on an **Intel 13700 CPU**. Platforms like Raspberry Pi, Orange Pi, and other ARM-based SBCs have different levels of processing capability, which may limit total FPS.
 
+### DEEPX NPU
+
+Frigate supports the DEEPX NPU in both of its form factors: the **DX-M1** M.2 module, which works on x86 (Intel/AMD) and ARM-based SBCs such as the Raspberry Pi 5, and the **DX-M1M** on the [Sixfab AI HAT+](https://docs.sixfab.com/docs/ai-hat-plus-raspberry-pi-5-quickstart) for the Raspberry Pi 5. Both use the same driver and runtime, so the configuration is identical for either one. DEEPX NPU support in Frigate is developed and maintained by [Sixfab](https://sixfab.com).
+
+The DEEPX driver and runtime run on the Docker host rather than inside the Frigate container and must be installed before the NPU can be used. See the [installation docs](installation.md#deepx-npu) for the setup steps and [the detector docs](/configuration/object_detectors#deepx-npu) for the configuration.
+
+Frigate does not bundle a model for this detector. Models use DEEPX's `.dxnn` format, and pre-compiled YOLO models can be downloaded from the [DEEPX ModelZoo](https://developer.deepx.ai/modelzoo). Prefer a model with a `_ppu` suffix whenever one is available for the architecture you want: these run part of the post-processing on the NPU itself and are considerably faster, roughly 2.5x for the same architecture and input size. **YOLOX-S with PPU is the recommended starting point.**
+
+Inference times for a few recommended models, measured through Frigate's own stats on a DX-M1:
+
+| Model             | Input Size | DX-M1 Inference Time |
+| ----------------- | ---------- | -------------------- |
+| YOLOX-S (PPU)     | 640        | ~ 13 ms              |
+| YOLOv9-t (PPU)    | 640        | ~ 18 ms              |
+| YOLOv4 (PPU)      | 512        | ~ 20 ms              |
+| YOLOX-S           | 640        | ~ 34 ms              |
+| YOLOv9-s          | 640        | ~ 39 ms              |
+
+Other ModelZoo YOLO variants are also supported but have not been measured. Inference speeds vary with the host platform, so a slower host such as a Raspberry Pi 5 will report higher times than those above.
+
+:::note
+
+A few ModelZoo models can not be used with Frigate: SSD models (they are trained on Pascal VOC, so their labels do not match Frigate's), DAMO-YOLO models, face and pose models, and the PPU builds of YOLOv7.
+
+:::
+
 ### Nvidia Jetson
 
 Jetson devices are supported via the TensorRT or ONNX detectors when running Jetpack 6. It will [make use of the Jetson's hardware media engine](/configuration/hardware_acceleration_video#nvidia-jetson) when configured with the [appropriate presets](/configuration/ffmpeg_presets#hwaccel-presets), and will make use of the Jetson's GPU and DLA for object detection when configured with the [TensorRT detector](/configuration/object_detectors#nvidia-tensorrt-detector).

docs/docs/frigate/installation.md

@@ -381,6 +381,99 @@ If you can't use Docker Compose, you can run the container with something simila
 
 Finally, configure [hardware object detection](/configuration/object_detectors#memryx-mx3) to complete the setup.
 
+### DEEPX NPU
+
+The DEEPX NPU is available in two form factors, and Frigate supports both:
+
+- **DX-M1** in the M.2 2280 form factor (like an NVMe SSD), for x86 (Intel/AMD) PCs, the Raspberry Pi 5, and other ARM SBCs with an exposed PCIe M.2 slot.
+- **DX-M1M** on the [Sixfab AI HAT+](https://docs.sixfab.com/docs/ai-hat-plus-raspberry-pi-5-quickstart), a HAT+ board that connects to the Raspberry Pi 5 over PCIe Gen 3 x1.
+
+Both present the NPU through the same PCIe driver and DX-RT runtime, so the setup below and the detector configuration are identical for either one. Nothing needs to change when moving between them.
+
+DEEPX NPU support in Frigate is developed and maintained by [Sixfab](https://sixfab.com).
+
+#### Versions
+
+A DEEPX install has several separately versioned pieces, and they all have to agree. The driver, the runtime, and the daemon live on the Docker host; Frigate itself carries only the Python bindings, which it downloads on first start:
+
+| Component      | Version  | Installed on | Installed by              |
+| -------------- | -------- | ------------ | ------------------------- |
+| Kernel driver  | `v2.6.0` | Host  | `user_installation.sh`    |
+| DX-RT runtime  | `v3.4.0` | Host  | `user_installation.sh`    |
+| NPU firmware   | `v2.7.4` | The module   | Flashed from the host     |
+| DX-RT bindings | `v3.4.0` | Frigate      | Downloaded at first start |
+
+:::warning
+
+A version mismatch does not produce a startup error. It typically shows up as inference requests that are accepted but never return a result, so detections simply stop appearing while Frigate looks healthy. If that happens after a Frigate upgrade, check every version in the table before anything else.
+
+:::
+
+The installation script installs the DX-RT runtime on the host and enables `dxrt.service`, so the daemon starts at boot and any other program on the host can share the NPU with Frigate. Check the firmware version with `dxrt-cli --status` and update the module if it does not match the table above.
+
+#### Installation
+
+The DEEPX kernel driver must be installed on the host rather than in the container, because containers share the host kernel and cannot load kernel modules. Installing it creates the `/dev/dxrt*` device nodes that are passed through to Frigate. The same script installs the DX-RT runtime and enables `dxrt.service`, the daemon that owns the NPU and hands work to it on behalf of Frigate and anything else on the host.
+
+1. Copy or download [this script](https://github.com/blakeblackshear/frigate/blob/dev/docker/deepx/user_installation.sh).
+2. Ensure it has execution permissions with `sudo chmod +x user_installation.sh`
+3. Run the script with `./user_installation.sh`
+4. **Restart your computer** to complete driver installation.
+
+Confirm the NPU is visible before continuing:
+
+```bash
+ls /dev/dxrt*
+```
+
+Then confirm the daemon is running and listening in `/run/dxrt`:
+
+```bash
+systemctl is-active dxrt.service
+ls /run/dxrt/
+```
+
+#### Setup
+
+To set up Frigate, follow the default installation instructions, for example: `ghcr.io/blakeblackshear/frigate:stable`
+
+#### Docker configuration
+
+Frigate needs the NPU device node and the directory holding the daemon's socket:
+
+```yaml
+services:
+  frigate:
+    devices:
+      - /dev/dxrt0:/dev/dxrt0
+    volumes:
+      - /run/dxrt:/run/dxrt
+```
+
+If you can't use Docker Compose, add `--device /dev/dxrt0:/dev/dxrt0 -v /run/dxrt:/run/dxrt` to your `docker run` command.
+
+Add one `--device` per NPU, contiguously from `/dev/dxrt0`, since the client stops enumerating at the first gap.
+
+The installation script configures `dxrt.service` to place its socket in `/run/dxrt` through a systemd drop-in. Mounting the directory rather than the socket file means the container sees the new socket after `dxrt.service` is restarted, rather than holding on to a deleted one.
+
+`dxrtd` listens on an abstract socket as well, but that one does not cross into a container, so Frigate names the filesystem socket through `DXRT_DYNAMIC_IPC_ENDPOINT` on your behalf. Set that variable on the container yourself only if the daemon listens somewhere else, which means you also set it for `dxrtd` through its own systemd drop-in. The script writes `/etc/systemd/system/dxrt.service.d/frigate.conf` for exactly that, and has `dxrt.service` link the socket to `/tmp/dxrt_dynamic_ipc.sock` when it starts, so the host's own `dxrt-cli` and `dxtop` keep finding it at the default path they fall back to.
+
+:::note
+
+The DX-RT client exits when `dxrt.service` stops, so restart the Frigate container after restarting `dxrt.service`.
+
+:::
+
+The device node is needed as well as the socket, because the client opens the NPU directly even though the daemon arbitrates access. Without it, inference fails with `Device not found`.
+
+`/dev/shm` does not need sharing.
+
+The DX-RT python bindings are not shipped in the Frigate image. Frigate downloads them on first start when a DEEPX detector is configured, and caches them under `/config`.
+
+#### Configuration
+
+Finally, configure [hardware object detection](/configuration/object_detectors#deepx-npu) to complete the setup.
+
 ### Rockchip platform
 
 Make sure that you use a linux distribution that comes with the rockchip BSP kernel 5.10 or 6.1 and necessary drivers (especially rkvdec2 and rknpu). To check, enter the following commands:

frigate/__main__.py

@@ -99,6 +99,10 @@ def main() -> None:
         print("***    End Config Validation Errors                       ***")
         print("*************************************************************")
 
+        # force a non-zero exit code for config failures
+        if args.validate_config:
+            sys.exit(1)
+
         # attempt to start Frigate in recovery mode
         try:
             config = FrigateConfig.load(install=True, safe_load=True)

frigate/api/media.py

@@ -258,15 +258,15 @@ async def latest_frame(
 
                 frame = request.app.camera_error_image
 
-        height = int(params.height or str(frame.shape[0]))
-        width = int(height * frame.shape[1] / frame.shape[0])
-
         if frame is None:
             return JSONResponse(
                 content={"success": False, "message": "Unable to get valid frame"},
                 status_code=500,
             )
 
+        height = int(params.height or str(frame.shape[0]))
+        width = int(height * frame.shape[1] / frame.shape[0])
+
         if height < 1 or width < 1:
             return JSONResponse(
                 content="Invalid height / width requested :: {} / {}".format(
@@ -886,9 +886,8 @@ async def vod_event(
     # If the recordings are not found and the event started more than 5 minutes ago, set has_clip to false
     if (
         event.start_time < datetime.now().timestamp() - 300
-        and type(vod_response) is tuple
-        and len(vod_response) == 2
-        and vod_response[1] == 404
+        and isinstance(vod_response, JSONResponse)
+        and vod_response.status_code == 404
     ):
         Event.update(has_clip=False).where(Event.id == event_id).execute()
 
@@ -956,64 +955,80 @@ async def event_snapshot(
     event_complete = False
     jpg_bytes = None
     frame_time = 0
+
     try:
         event = Event.get(Event.id == event_id, Event.end_time != None)
-        event_complete = True
         await require_camera_access(event.camera, request=request)
+    except DoesNotExist:
+        event = None
+
+    if event is not None:
+        event_complete = True
+
         if not event.has_snapshot:
             return JSONResponse(
                 content={"success": False, "message": "Snapshot not available"},
                 status_code=404,
             )
-        snapshot_settings = _resolve_snapshot_settings(
-            request.app.frigate_config.cameras[event.camera].snapshots, params
-        )
-        jpg_bytes, frame_time = get_event_snapshot_bytes(
-            event,
-            ext="jpg",
-            timestamp=snapshot_settings["timestamp"],
-            bounding_box=snapshot_settings["bounding_box"],
-            crop=snapshot_settings["crop"],
-            height=snapshot_settings["height"],
-            quality=snapshot_settings["quality"],
-            timestamp_style=request.app.frigate_config.cameras[
-                event.camera
-            ].timestamp_style,
-            colormap=request.app.frigate_config.model_for_camera(event.camera).colormap,
-        )
-    except DoesNotExist:
-        # see if the object is currently being tracked
+
         try:
-            camera_states: list[CameraState] = (
-                request.app.detected_frames_processor.get_camera_states()
+            snapshot_settings = _resolve_snapshot_settings(
+                request.app.frigate_config.cameras[event.camera].snapshots, params
+            )
+            jpg_bytes, frame_time = get_event_snapshot_bytes(
+                event,
+                ext="jpg",
+                timestamp=snapshot_settings["timestamp"],
+                bounding_box=snapshot_settings["bounding_box"],
+                crop=snapshot_settings["crop"],
+                height=snapshot_settings["height"],
+                quality=snapshot_settings["quality"],
+                timestamp_style=request.app.frigate_config.cameras[
+                    event.camera
+                ].timestamp_style,
+                colormap=request.app.frigate_config.model_for_camera(
+                    event.camera
+                ).colormap,
             )
-            for camera_state in camera_states:
-                if event_id in camera_state.tracked_objects:
-                    tracked_obj = camera_state.tracked_objects.get(event_id)
-                    if tracked_obj is not None:
-                        snapshot_settings = _resolve_snapshot_settings(
-                            camera_state.camera_config.snapshots, params
-                        )
-                        jpg_bytes, frame_time = tracked_obj.get_img_bytes(
-                            ext="jpg",
-                            timestamp=snapshot_settings["timestamp"],
-                            bounding_box=snapshot_settings["bounding_box"],
-                            crop=snapshot_settings["crop"],
-                            height=snapshot_settings["height"],
-                            quality=snapshot_settings["quality"],
-                        )
-                        await require_camera_access(camera_state.name, request=request)
         except Exception:
             return JSONResponse(
-                content={"success": False, "message": "Ongoing event not found"},
+                content={"success": False, "message": "Unknown error occurred"},
                 status_code=404,
             )
-    except Exception:
-        return JSONResponse(
-            content={"success": False, "message": "Unknown error occurred"},
-            status_code=404,
+    else:
+        # see if the object is currently being tracked
+        camera_states: list[CameraState] = (
+            request.app.detected_frames_processor.get_camera_states()
         )
 
+        for camera_state in camera_states:
+            tracked_obj = camera_state.tracked_objects.get(event_id)
+
+            if tracked_obj is None:
+                continue
+
+            await require_camera_access(camera_state.name, request=request)
+
+            try:
+                snapshot_settings = _resolve_snapshot_settings(
+                    camera_state.camera_config.snapshots, params
+                )
+                jpg_bytes, frame_time = tracked_obj.get_img_bytes(
+                    ext="jpg",
+                    timestamp=snapshot_settings["timestamp"],
+                    bounding_box=snapshot_settings["bounding_box"],
+                    crop=snapshot_settings["crop"],
+                    height=snapshot_settings["height"],
+                    quality=snapshot_settings["quality"],
+                )
+            except Exception:
+                return JSONResponse(
+                    content={"success": False, "message": "Ongoing event not found"},
+                    status_code=404,
+                )
+
+            break
+
     if jpg_bytes is None:
         return JSONResponse(
             content={"success": False, "message": "Live frame not available"},
@@ -1062,19 +1077,25 @@ async def event_thumbnail(
 
     if not thumbnail_bytes:
         # see if the object is currently being tracked
-        try:
-            camera_states = request.app.detected_frames_processor.get_camera_states()
-            for camera_state in camera_states:
-                if event_id in camera_state.tracked_objects:
-                    tracked_obj = camera_state.tracked_objects.get(event_id)
-                    if tracked_obj is not None:
-                        await require_camera_access(camera_state.name, request=request)
-                        thumbnail_bytes = tracked_obj.get_thumbnail(extension.value)
-        except Exception:
-            return JSONResponse(
-                content={"success": False, "message": "Event not found"},
-                status_code=404,
-            )
+        camera_states = request.app.detected_frames_processor.get_camera_states()
+
+        for camera_state in camera_states:
+            tracked_obj = camera_state.tracked_objects.get(event_id)
+
+            if tracked_obj is None:
+                continue
+
+            await require_camera_access(camera_state.name, request=request)
+
+            try:
+                thumbnail_bytes = tracked_obj.get_thumbnail(extension.value)
+            except Exception:
+                return JSONResponse(
+                    content={"success": False, "message": "Event not found"},
+                    status_code=404,
+                )
+
+            break
 
     if not thumbnail_bytes:
         return JSONResponse(

frigate/app.py

@@ -483,7 +483,7 @@ class FrigateApp:
 
     def start_audio_processor(self) -> None:
         self.audio_process = AudioProcessor(
-            self.config, self.camera_metrics, self.stop_event
+            self.config, self.camera_metrics, self.embeddings_metrics, self.stop_event
         )
         self.audio_process.start()
         self.processes["audio_detector"] = self.audio_process.pid or 0
@@ -556,10 +556,20 @@ class FrigateApp:
                 "output",
                 lambda: OutputProcess(self.config, self.stop_event),
             ),
+            (
+                "audio_process",
+                "audio_detector",
+                lambda: AudioProcessor(
+                    self.config,
+                    self.camera_metrics,
+                    self.embeddings_metrics,
+                    self.stop_event,
+                ),
+            ),
         ]
 
         for attr, key, factory in specs:
-            if not hasattr(self, attr):
+            if getattr(self, attr, None) is None:
                 continue
 
             def on_restart(

frigate/config/camera/genai.py

@@ -1,7 +1,7 @@
 from enum import Enum
-from typing import Any
+from typing import Any, Self
 
-from pydantic import Field
+from pydantic import Field, model_validator
 
 from ..base import FrigateBaseModel
 from ..env import EnvString
@@ -21,6 +21,17 @@ class GenAIRoleEnum(str, Enum):
     chat = "chat"
     descriptions = "descriptions"
     embeddings = "embeddings"
+    transcribe = "transcribe"
+
+
+# Providers that can accept audio input for the transcribe role. Ollama has no
+# audio input support, so claiming the role there would fail at request time.
+TRANSCRIBE_CAPABLE_PROVIDERS = {
+    GenAIProviderEnum.openai,
+    GenAIProviderEnum.azure_openai,
+    GenAIProviderEnum.gemini,
+    GenAIProviderEnum.llamacpp,
+}
 
 
 class GenAIConfig(FrigateBaseModel):
@@ -52,7 +63,7 @@ class GenAIConfig(FrigateBaseModel):
             GenAIRoleEnum.chat,
         ],
         title="Roles",
-        description="GenAI roles (chat, descriptions, embeddings); one provider per role.",
+        description="GenAI roles (chat, descriptions, embeddings, transcribe); one provider per role. Only chat, descriptions, and embeddings are granted by default; transcribe must be listed explicitly.",
     )
     provider_options: dict[str, Any] = Field(
         default={},
@@ -66,3 +77,17 @@ class GenAIConfig(FrigateBaseModel):
         description="Runtime options passed to the provider for each inference call.",
         json_schema_extra={"additionalProperties": {}},
     )
+
+    @model_validator(mode="after")
+    def validate_transcribe_provider(self) -> Self:
+        """Reject the transcribe role on providers that cannot accept audio input."""
+        if (
+            GenAIRoleEnum.transcribe in self.roles
+            and self.provider not in TRANSCRIBE_CAPABLE_PROVIDERS
+        ):
+            raise ValueError(
+                f"GenAI provider '{self.provider.value}' does not support audio input "
+                "and cannot be given the 'transcribe' role."
+            )
+
+        return self

frigate/config/camera/review.py

@@ -9,6 +9,7 @@ __all__ = [
     "DetectionsConfig",
     "AlertsConfig",
     "ImageSourceEnum",
+    "ReviewFrameModeEnum",
     "ReviewResponseStyleEnum",
 ]
 
@@ -20,6 +21,13 @@ class ImageSourceEnum(str, Enum):
     recordings = "recordings"
 
 
+class ReviewFrameModeEnum(str, Enum):
+    """How review frames are presented to the GenAI provider."""
+
+    frames = "frames"
+    annotated_frames = "annotated_frames"
+
+
 class ReviewResponseStyleEnum(str, Enum):
     """Writing style presets for GenAI review descriptions."""
 
@@ -153,6 +161,11 @@ class GenAIReviewConfig(FrigateBaseModel):
         description="Preferred language to request from the GenAI provider for generated responses.",
         default=None,
     )
+    frame_mode: ReviewFrameModeEnum = Field(
+        default=ReviewFrameModeEnum.frames,
+        title="Frame mode",
+        description="How frames are presented to the model. 'frames' sends the prompt followed by the frames, which suits models that track a sequence well on their own. 'annotated_frames' labels each frame and interleaves notes derived from object tracking, which helps models that lose track of activity that repeats or reverses.",
+    )
     response_style: ReviewResponseStyleEnum = Field(
         default=ReviewResponseStyleEnum.default,
         title="Response style",

frigate/config/classification.py

@@ -5,6 +5,7 @@ from pydantic import ConfigDict, Field, field_validator
 from .base import FrigateBaseModel
 
 __all__ = [
+    "AudioTranscriptionModelEnum",
     "CameraFaceRecognitionConfig",
     "CameraLicensePlateRecognitionConfig",
     "CameraAudioTranscriptionConfig",
@@ -20,6 +21,10 @@ class SemanticSearchModelEnum(str, Enum):
     jinav2 = "jinav2"
 
 
+class AudioTranscriptionModelEnum(str, Enum):
+    whisper = "whisper"
+
+
 class EnrichmentsDeviceEnum(str, Enum):
     GPU = "GPU"
     CPU = "CPU"
@@ -53,10 +58,35 @@ class AudioTranscriptionConfig(FrigateBaseModel):
         description="Enable or disable automatic audio transcription for all cameras; can be overridden per-camera.",
     )
     language: str = Field(
-        default="en",
+        default="auto",
         title="Transcription language",
-        description="Language code used for transcription/translation (for example 'en' for English). See https://whisper-api.com/docs/languages/ for supported language codes.",
+        description="Language code used for transcription/translation (for example 'en' for English), or 'auto' to let the model detect it. See https://whisper-api.com/docs/languages/ for supported language codes.",
     )
+    model: AudioTranscriptionModelEnum | str | None = Field(
+        default=AudioTranscriptionModelEnum.whisper,
+        title="Audio transcription model or GenAI provider name",
+        description="The transcription backend: 'whisper' for Frigate's built-in local models, or the name of a GenAI provider with the transcribe role.",
+    )
+
+    @field_validator("model", mode="before")
+    @classmethod
+    def coerce_model_enum(cls, v):
+        # An absent value ("model:" with nothing after it, or an explicit null)
+        # means unspecified, so fall back to the built-in backend. Left as None
+        # it would pass the GenAI-provider validation, which only inspects
+        # strings, and then be treated as a provider name that resolves to no
+        # client, turning transcription into a silent no-op.
+        if v is None or (isinstance(v, str) and not v.strip()):
+            return AudioTranscriptionModelEnum.whisper
+
+        if isinstance(v, str):
+            try:
+                return AudioTranscriptionModelEnum(v)
+            except ValueError:
+                return v
+
+        return v
+
     device: EnrichmentsDeviceEnum = Field(
         default=EnrichmentsDeviceEnum.CPU,
         title="Transcription device",

frigate/config/config.py

@@ -56,6 +56,7 @@ from .camera.timestamp import TimestampStyleConfig
 from .camera_group import CameraGroupConfig
 from .classification import (
     AudioTranscriptionConfig,
+    AudioTranscriptionModelEnum,
     ClassificationConfig,
     FaceRecognitionConfig,
     LicensePlateRecognitionConfig,
@@ -884,7 +885,7 @@ class FrigateConfig(FrigateBaseModel):
         # set notifications state
         self.notifications.enabled_in_config = self.notifications.enabled
 
-        # validate genai: each role (chat, descriptions, embeddings) at most once
+        # validate genai: each role (chat, descriptions, embeddings, transcribe) at most once
         role_to_name: dict[GenAIRoleEnum, str] = {}
         for name, genai_cfg in self.genai.items():
             for role in genai_cfg.roles:
@@ -1245,6 +1246,35 @@ class FrigateConfig(FrigateBaseModel):
         for model in self.models:
             model.create_colormap(colored_labels)
 
+        # validate audio_transcription.model when it is a GenAI provider name.
+        # this runs here rather than beside the semantic_search check because the
+        # global->camera merge above is what resolves camera-level enablement.
+        transcription_active = self.audio_transcription.enabled or any(
+            camera.audio_transcription.enabled for camera in self.cameras.values()
+        )
+
+        if (
+            transcription_active
+            and isinstance(self.audio_transcription.model, str)
+            and not isinstance(
+                self.audio_transcription.model, AudioTranscriptionModelEnum
+            )
+        ):
+            if self.audio_transcription.model not in self.genai:
+                raise ValueError(
+                    f"audio_transcription.model '{self.audio_transcription.model}' is not a "
+                    "valid GenAI config key. Must match a key in genai config."
+                )
+
+            if (
+                GenAIRoleEnum.transcribe
+                not in self.genai[self.audio_transcription.model].roles
+            ):
+                raise ValueError(
+                    f"GenAI provider '{self.audio_transcription.model}' must have "
+                    "'transcribe' in its roles for audio transcription."
+                )
+
         # Check audio transcription and audio detection requirements
         if self.audio_transcription.enabled:
             # If audio transcription is enabled globally, at least one camera must have audio detection enabled

frigate/data_processing/post/audio_transcription.py

@@ -10,6 +10,7 @@ from peewee import DoesNotExist
 
 from frigate.comms.inter_process import InterProcessRequestor
 from frigate.config import FrigateConfig
+from frigate.config.classification import AudioTranscriptionModelEnum
 from frigate.const import (
     CACHE_DIR,
     MODEL_CACHE_DIR,
@@ -18,8 +19,13 @@ from frigate.const import (
 )
 from frigate.data_processing.types import PostProcessDataEnum
 from frigate.embeddings.embeddings import Embeddings
+from frigate.genai.manager import GenAIClientManager
 from frigate.types import TrackedObjectUpdateTypesEnum
-from frigate.util.audio import get_audio_from_recording
+from frigate.util.audio import (
+    clean_transcript,
+    get_audio_from_recording,
+    resolve_language,
+)
 
 from ..types import DataProcessorMetrics
 from .api import PostProcessorApi
@@ -34,15 +40,25 @@ class AudioTranscriptionPostProcessor(PostProcessorApi):
         requestor: InterProcessRequestor,
         embeddings: Embeddings,
         metrics: DataProcessorMetrics,
+        genai_manager: GenAIClientManager | None = None,
     ):
         super().__init__(config, metrics, None)
         self.config = config
         self.requestor = requestor
         self.embeddings = embeddings
+        self.genai_manager = genai_manager
         self.recognizer = None
         self.transcription_lock = threading.Lock()
         self.transcription_thread: threading.Thread | None = None
         self.transcription_running = False
+        self._use_genai = not isinstance(
+            config.audio_transcription.model, AudioTranscriptionModelEnum
+        )
+
+        if self._use_genai:
+            # never build the local recognizer on the GenAI path; WhisperModel
+            # downloads several hundred MB on first use
+            return
 
         # faster-whisper handles model downloading automatically
         self.model_path = os.path.join(MODEL_CACHE_DIR, "whisper")
@@ -147,6 +163,31 @@ class AudioTranscriptionPostProcessor(PostProcessorApi):
             logger.error(f"Error in audio transcription post-processing: {e}")
 
     def __transcribe_audio(self, audio_data: bytes) -> str | None:
+        """Transcribe WAV audio data with the configured backend."""
+        if self._use_genai:
+            return self.__transcribe_audio_genai(audio_data)
+
+        return self.__transcribe_audio_whisper(audio_data)
+
+    def __transcribe_audio_genai(self, audio_data: bytes) -> str | None:
+        """Hand the WAV bytes to the GenAI provider holding the transcribe role."""
+        client = self.genai_manager.transcribe_client if self.genai_manager else None
+
+        if not client:
+            logger.error(
+                "audio_transcription.model is '%s' (GenAI provider) but no transcribe "
+                "client is configured. Ensure the GenAI provider has 'transcribe' in its roles",
+                self.config.audio_transcription.model,
+            )
+            return None
+
+        text = client.transcribe(
+            audio_data,
+            language=resolve_language(self.config.audio_transcription.language),
+        )
+        return clean_transcript(text) or None
+
+    def __transcribe_audio_whisper(self, audio_data: bytes) -> str | None:
         """Transcribe WAV audio data using faster-whisper."""
         if not self.recognizer:
             logger.debug("Recognizer not initialized")
@@ -160,7 +201,7 @@ class AudioTranscriptionPostProcessor(PostProcessorApi):
 
             segments, info = self.recognizer.transcribe(
                 temp_wav,
-                language=self.config.audio_transcription.language,
+                language=resolve_language(self.config.audio_transcription.language),
                 beam_size=5,
             )
 

frigate/data_processing/post/review_annotations.py

@@ -0,0 +1,394 @@
+"""Frame annotations derived from object tracking data.
+
+Builds short notes describing what changed during a review item, keyed to the
+frames sampled from it. Everything here comes from tracked object data already
+in the database (each event's `path_data` trajectory and the timeline's
+stationary/active changes), so the notes can be stated to the model as fact
+rather than as something it must perceive.
+"""
+
+import logging
+import math
+from collections.abc import Sequence
+from typing import Any
+
+from frigate.models import Event, Timeline
+
+logger = logging.getLogger(__name__)
+
+# Movement smaller than this (normalized frame units) between two path points
+# is treated as the object holding still rather than travelling.
+STILL_THRESHOLD = 0.02
+
+# A heading change beyond this (dot product against the leg's own heading)
+# counts as the object turning back rather than curving.
+REVERSAL_DOT = -0.3
+
+# A run of travel shorter than this (normalized frame units) is treated as
+# milling about rather than going somewhere. Without it, a subject pacing in
+# one spot produces a burst of contradictory "turns around" notes on a single
+# frame.
+MIN_LEG_DISTANCE = 0.08
+
+# Movement that begins within this many seconds of detection is folded into
+# the detection note, so each arrival reads as one event instead of several.
+DETECT_MOVE_MERGE_SECONDS = 2.0
+
+STATE_CHANGE_PHRASES = {
+    "stationary": "has stopped moving",
+    "active": "starts moving again",
+}
+
+Point = tuple[float, float, float]
+Leg = tuple[int, int]
+
+
+def describe_position(x: float, y: float) -> str:
+    """Name a normalized frame position in plain terms."""
+    horizontal = "left" if x < 0.34 else ("right" if x > 0.66 else "center")
+    vertical = "top" if y < 0.34 else ("bottom" if y > 0.66 else "middle")
+
+    if horizontal == "center" and vertical == "middle":
+        return "the middle of the frame"
+
+    if horizontal == "center":
+        return f"the {vertical} of the frame"
+
+    if vertical == "middle":
+        return f"the {horizontal} of the frame"
+
+    return f"the {vertical} {horizontal} of the frame"
+
+
+def describe_heading(dx: float, dy: float) -> str:
+    """Name a direction of travel in frame terms.
+
+    y grows downward in normalized coordinates, so a falling y reads as moving
+    toward the top of the frame.
+    """
+    parts = []
+
+    if abs(dy) > abs(dx) * 0.4:
+        parts.append("down" if dy > 0 else "up")
+
+    if abs(dx) > abs(dy) * 0.4:
+        parts.append("right" if dx > 0 else "left")
+
+    return " and ".join(parts) if parts else "in place"
+
+
+def event_name(event: dict[str, Any]) -> str:
+    """Name an object for the notes, e.g. 'a person' or 'waste bin "Compost"'.
+
+    Objects are never numbered or given track identifiers. Frigate opens a new
+    tracked object whenever a subject is re-detected, so the tracking data
+    cannot say whether two entries are the same subject, and the notes stay
+    ambiguous rather than implying either answer.
+    """
+    label = str(event["label"]).replace("_", " ").replace("-verified", "")
+    sub_label = event.get("sub_label")
+
+    if sub_label:
+        return f'{label} "{sub_label}"'
+
+    article = "an" if label[:1].lower() in "aeiou" else "a"
+    return f"{article} {label}"
+
+
+def path_legs(points: list[Point]) -> list[Leg]:
+    """Split a trajectory into runs of travel in a consistent direction.
+
+    A leg ends when the subject starts moving back against the direction that
+    leg established, and only once the leg has covered MIN_LEG_DISTANCE, so
+    jitter around a standing subject does not register as a turn.
+
+    Returns (start, end) index pairs into `points`.
+    """
+    legs: list[Leg] = []
+    start = 0
+
+    for i in range(1, len(points)):
+        lx = points[i][0] - points[start][0]
+        ly = points[i][1] - points[start][1]
+        leg_distance = (lx * lx + ly * ly) ** 0.5
+
+        if leg_distance < MIN_LEG_DISTANCE:
+            continue
+
+        sx = points[i][0] - points[i - 1][0]
+        sy = points[i][1] - points[i - 1][1]
+        step = (sx * sx + sy * sy) ** 0.5
+
+        if step < STILL_THRESHOLD:
+            continue
+
+        dot = (lx / leg_distance) * (sx / step) + (ly / leg_distance) * (sy / step)
+
+        if dot < REVERSAL_DOT:
+            legs.append((start, i - 1))
+            start = i - 1
+
+    if start < len(points) - 1:
+        legs.append((start, len(points) - 1))
+
+    return [
+        (a, b)
+        for a, b in legs
+        if ((points[b][0] - points[a][0]) ** 2 + (points[b][1] - points[a][1]) ** 2)
+        ** 0.5
+        >= MIN_LEG_DISTANCE
+    ]
+
+
+def path_points(path_data: list[Any]) -> list[Point]:
+    """Flatten path_data into (x, y, timestamp) tuples, or [] if malformed."""
+    try:
+        return [(p[0][0], p[0][1], p[1]) for p in path_data or []]
+    except (IndexError, TypeError):
+        logger.debug("Malformed path_data, skipping trajectory notes")
+        return []
+
+
+def leg_start_time(points: list[Point], leg: Leg) -> float:
+    """When a leg's movement actually began.
+
+    path_data always keeps an object's first two samples, so a leg can open
+    with points recorded long before the object moved. The first sample that
+    has left the leg's origin is the earliest evidence of movement.
+    """
+    a, b = leg
+    x0, y0, t0 = points[a]
+
+    for x, y, t in points[a + 1 : b + 1]:
+        if math.hypot(x - x0, y - y0) >= STILL_THRESHOLD:
+            return t
+
+    return t0
+
+
+def leg_heading(points: list[Point], leg: Leg) -> str:
+    a, b = leg
+    return describe_heading(points[b][0] - points[a][0], points[b][1] - points[a][1])
+
+
+def leg_phrase(points: list[Point], leg: Leg, first: bool) -> str:
+    """Describe the start of a leg, e.g. 'turns around at ... and heads left'."""
+    heading = leg_heading(points, leg)
+    place = describe_position(points[leg[0]][0], points[leg[0]][1])
+
+    if first:
+        return f"starts moving {heading} from {place}"
+
+    return f"turns around at {place} and heads {heading}"
+
+
+def path_moments(path_data: list[Any]) -> list[tuple[float, str]]:
+    """Key moments in one trajectory as (timestamp, phrase).
+
+    Emits one note per leg of travel. Where the last leg ends is left out:
+    path_data only records significant movement, so its final point cannot
+    distinguish an object coming to rest from one leaving the frame.
+    """
+    points = path_points(path_data)
+
+    if len(points) < 2:
+        return []
+
+    return [
+        (leg_start_time(points, leg), leg_phrase(points, leg, index == 0))
+        for index, leg in enumerate(path_legs(points))
+    ]
+
+
+def build_timeline(
+    events: list[dict[str, Any]],
+    span_end: float,
+    state_changes: Sequence[dict[str, Any]] = (),
+) -> list[tuple[float, str]]:
+    """All annotated moments across every event, in time order.
+
+    Only changes are noted, since those are what sparse frames miss; an
+    object's state at the end of the clip is visible in the last frame.
+    `span_end` is the timestamp of the last sampled frame, and moments past it
+    describe nothing the model can see. A track ending means the object
+    stopped being detected, which may or may not mean it left the frame.
+
+    `state_changes` are timeline rows (timestamp, source_id, class_type); the
+    stationary and active ones become "has stopped moving" / "starts moving
+    again". Frigate only marks an object stationary after it has been still
+    for a while, which the past-tense wording reflects.
+
+    Each object keeps its own notes. Folding an object into the note of the
+    person moving it ("alongside ...") was tried and made models lose track of
+    where the object went.
+    """
+    changes_by_event: dict[str, list[tuple[float, str]]] = {}
+
+    for change in state_changes:
+        phrase = STATE_CHANGE_PHRASES.get(change["class_type"])
+
+        if phrase:
+            changes_by_event.setdefault(change["source_id"], []).append(
+                (change["timestamp"], phrase)
+            )
+
+    timeline: list[tuple[float, str]] = []
+
+    for event in sorted(events, key=lambda e: e["start_time"]):
+        # Frame extraction can come up short at the end of a clip, leaving
+        # objects that only appear after the last frame we actually have.
+        if event["start_time"] > span_end:
+            continue
+
+        points = path_points(event.get("path_data") or [])
+        legs = path_legs(points) if len(points) >= 2 else []
+        name = event_name(event)
+        detected_at = event["start_time"]
+        where = describe_position(points[0][0], points[0][1]) if points else "the frame"
+        merges = bool(legs) and leg_start_time(points, legs[0]) - detected_at <= (
+            DETECT_MOVE_MERGE_SECONDS
+        )
+        remaining = list(enumerate(legs))
+        moments: list[tuple[float, str]] = []
+
+        if merges:
+            heading = leg_heading(points, legs[0])
+            timeline.append(
+                (detected_at, f"{name} first detected at {where}, moving {heading}")
+            )
+            remaining = remaining[1:]
+        else:
+            timeline.append((detected_at, f"{name} first detected at {where}"))
+
+        for index, leg in remaining:
+            moments.append(
+                (
+                    leg_start_time(points, leg),
+                    leg_phrase(points, leg, index == 0),
+                )
+            )
+
+        moments.extend(
+            (timestamp, phrase)
+            for timestamp, phrase in changes_by_event.get(event["id"], [])
+            if timestamp >= detected_at
+        )
+
+        for timestamp, phrase in moments:
+            if timestamp <= span_end:
+                timeline.append((timestamp, f"{name} {phrase}"))
+
+        if event["end_time"] and event["end_time"] <= span_end:
+            timeline.append((event["end_time"], f"{name} is no longer detected"))
+
+    return sorted(timeline, key=lambda m: m[0])
+
+
+def annotations_by_frame(
+    timeline: list[tuple[float, str]], frame_times: list[float]
+) -> dict[int, list[str]]:
+    """Bucket timeline moments onto the frame that follows each one.
+
+    A moment is attached to the first frame at or after it happened, so the
+    note always precedes the image in which the change becomes visible.
+    """
+    buckets: dict[int, list[str]] = {}
+
+    if not frame_times:
+        return buckets
+
+    for timestamp, phrase in timeline:
+        index = next(
+            (i for i, ft in enumerate(frame_times) if ft >= timestamp),
+            len(frame_times) - 1,
+        )
+        buckets.setdefault(index, []).append(phrase)
+
+    return buckets
+
+
+def get_tracked_events(detection_ids: list[str]) -> list[dict[str, Any]]:
+    """Load the tracked objects behind a review item's detections."""
+    if not detection_ids:
+        return []
+
+    rows = list(
+        Event.select(
+            Event.id,
+            Event.label,
+            Event.sub_label,
+            Event.start_time,
+            Event.end_time,
+            Event.data,
+        )
+        .where(Event.id << detection_ids)
+        .dicts()
+        .iterator()
+    )
+
+    return [
+        {
+            "id": row["id"],
+            "label": row["label"],
+            "sub_label": row["sub_label"],
+            "start_time": row["start_time"],
+            "end_time": row["end_time"],
+            "path_data": (row["data"] or {}).get("path_data") or [],
+        }
+        for row in rows
+        if row["start_time"] is not None
+    ]
+
+
+def get_state_changes(detection_ids: list[str]) -> list[dict[str, Any]]:
+    """Stationary/active changes the timeline recorded for these objects."""
+    if not detection_ids:
+        return []
+
+    return list(
+        Timeline.select(Timeline.timestamp, Timeline.source_id, Timeline.class_type)
+        .where(
+            (Timeline.source_id << detection_ids)
+            & (Timeline.class_type << list(STATE_CHANGE_PHRASES))
+        )
+        .dicts()
+        .iterator()
+    )
+
+
+def build_frame_captions(
+    detection_ids: list[str],
+    frame_times: list[float],
+) -> list[str]:
+    """A caption for each sampled frame, in frame order.
+
+    Every frame gets its index and elapsed time so the model can tell them
+    apart; frames where something changed also carry the tracker notes for
+    that moment. Returns an empty list when there is nothing to say, which
+    callers treat as a reason to fall back to sending plain frames.
+    """
+    if not frame_times:
+        return []
+
+    events = get_tracked_events(detection_ids)
+
+    if not events:
+        logger.debug("No tracked events found for review item, skipping annotations")
+        return []
+
+    timeline = build_timeline(events, frame_times[-1], get_state_changes(detection_ids))
+    buckets = annotations_by_frame(timeline, frame_times)
+
+    if not buckets:
+        return []
+
+    total = len(frame_times)
+    origin = frame_times[0]
+    captions: list[str] = []
+
+    for index, timestamp in enumerate(frame_times):
+        lines = [f"Frame {index + 1} of {total} (+{timestamp - origin:.1f}s):"]
+        lines.extend(f"[tracker] {note}" for note in buckets.get(index, []))
+        captions.append("\n".join(lines))
+
+    return captions

frigate/data_processing/post/review_descriptions.py

@@ -19,7 +19,11 @@ from frigate.comms.embeddings_updater import EmbeddingsRequestEnum
 from frigate.comms.inter_process import InterProcessRequestor
 from frigate.config import FrigateConfig
 from frigate.config.camera import CameraConfig
-from frigate.config.camera.review import GenAIReviewConfig, ImageSourceEnum
+from frigate.config.camera.review import (
+    GenAIReviewConfig,
+    ImageSourceEnum,
+    ReviewFrameModeEnum,
+)
 from frigate.const import (
     ATTRIBUTE_LABEL_DISPLAY_MAP,
     CACHE_DIR,
@@ -36,6 +40,7 @@ from frigate.util.image import get_image_from_recording
 
 from ..post.api import PostProcessorApi
 from ..types import DataProcessorMetrics
+from .review_annotations import build_frame_captions
 
 logger = logging.getLogger(__name__)
 
@@ -43,6 +48,7 @@ RECORDING_BUFFER_EXTENSION_PERCENT = 0.10
 MIN_RECORDING_DURATION = 10
 MAX_IMAGE_TOKENS = 24000
 MAX_FRAMES_PER_SECOND = 1
+MAX_ANNOTATED_FRAMES = 28
 
 
 class ReviewDescriptionProcessor(PostProcessorApi):
@@ -67,6 +73,7 @@ class ReviewDescriptionProcessor(PostProcessorApi):
         duration: float,
         image_source: ImageSourceEnum = ImageSourceEnum.preview,
         height: int = 480,
+        frame_mode: ReviewFrameModeEnum = ReviewFrameModeEnum.frames,
     ) -> int:
         """Calculate optimal number of frames based on event duration, context size,
         image source, and resolution.
@@ -80,6 +87,8 @@ class ReviewDescriptionProcessor(PostProcessorApi):
           - MAX_FRAMES_PER_SECOND x duration, to avoid drowning short events in
             near-duplicate frames where the model latches onto the redundant middle
             and skips the start/end action
+          - MAX_ANNOTATED_FRAMES in annotated mode, where the tracking notes
+            already carry the sequence
         """
         client = self.genai_manager.description_client
 
@@ -125,6 +134,10 @@ class ReviewDescriptionProcessor(PostProcessorApi):
         max_frames_by_tokens = int(image_token_budget / tokens_per_image)
         max_frames_by_duration = int(duration * MAX_FRAMES_PER_SECOND)
         max_frames = min(max_frames_by_tokens, max_frames_by_duration)
+
+        if frame_mode == ReviewFrameModeEnum.annotated_frames:
+            max_frames = min(max_frames, MAX_ANNOTATED_FRAMES)
+
         return max(max_frames, 3)
 
     def process_data(
@@ -166,6 +179,7 @@ class ReviewDescriptionProcessor(PostProcessorApi):
                 return
 
             image_source = camera_config.review.genai.image_source
+            frame_mode = camera_config.review.genai.frame_mode
 
             if image_source == ImageSourceEnum.recordings:
                 buffer_extension = get_recording_buffer_extension(
@@ -174,40 +188,43 @@ class ReviewDescriptionProcessor(PostProcessorApi):
                 final_data["start_time"] -= buffer_extension
                 final_data["end_time"] += buffer_extension
 
-                thumbs = self.get_recording_frames(
+                frames = self.get_recording_frames(
                     camera,
                     final_data["start_time"],
                     final_data["end_time"],
                     height=480,  # Use 480p for good balance between quality and token usage
+                    frame_mode=frame_mode,
                 )
 
-                if not thumbs:
+                if not frames:
                     # Fallback to preview frames if no recordings available
                     logger.warning(
                         f"No recording frames found for {camera}, falling back to preview frames"
                     )
-                    thumbs = self.get_preview_frames_as_bytes(
+                    frames = self.get_preview_frames_as_bytes(
                         camera,
                         final_data["start_time"],
                         final_data["end_time"],
                         final_data["thumb_path"],
                         id,
                         camera_config.review.genai.debug_save_thumbnails,
+                        frame_mode,
                     )
                 elif camera_config.review.genai.debug_save_thumbnails:
-                    self.save_debug_recording_frames(id, thumbs)
+                    self.save_debug_recording_frames(id, frames)
             else:
                 # Use preview frames
-                thumbs = self.get_preview_frames_as_bytes(
+                frames = self.get_preview_frames_as_bytes(
                     camera,
                     final_data["start_time"],
                     final_data["end_time"],
                     final_data["thumb_path"],
                     id,
                     camera_config.review.genai.debug_save_thumbnails,
+                    frame_mode,
                 )
 
-            self.start_analysis(camera_config, final_data, thumbs)
+            self.start_analysis(camera_config, final_data, frames)
 
     def handle_request(self, topic: str, request_data: dict[str, Any]) -> str | None:
         if topic == EmbeddingsRequestEnum.regenerate_review_description.value:
@@ -386,31 +403,50 @@ class ReviewDescriptionProcessor(PostProcessorApi):
         buffer_extension = get_recording_buffer_extension(
             final_data["end_time"] - final_data["start_time"]
         )
-        thumbs = self.get_recording_frames(
+        frames = self.get_recording_frames(
             str(review.camera),
             final_data["start_time"] - buffer_extension,
             final_data["end_time"] + buffer_extension,
             height=480,
+            frame_mode=camera_config.review.genai.frame_mode,
         )
 
-        if not thumbs:
+        if not frames:
             logger.error(
                 "No recording frames are available for review item %s", review_id
             )
             return
 
         if camera_config.review.genai.debug_save_thumbnails:
-            self.save_debug_recording_frames(review_id, thumbs)
+            self.save_debug_recording_frames(review_id, frames)
 
-        self.start_analysis(camera_config, final_data, thumbs)
+        self.start_analysis(camera_config, final_data, frames)
 
     def start_analysis(
         self,
         camera_config: CameraConfig,
         final_data: dict[str, Any],
-        thumbs: list[bytes],
+        frames: list[tuple[bytes, float]],
     ) -> None:
         """Kick off description generation for a review item in the background."""
+        thumbs = [frame for frame, _ in frames]
+        captions: list[str] = []
+
+        if (
+            camera_config.review.genai.frame_mode
+            == ReviewFrameModeEnum.annotated_frames
+        ):
+            captions = build_frame_captions(
+                final_data["data"].get("detections") or [],
+                [timestamp for _, timestamp in frames],
+            )
+
+            if not captions:
+                logger.debug(
+                    "No tracking annotations for review item %s, sending plain frames",
+                    final_data["id"],
+                )
+
         self.review_desc_dps.update()
         threading.Thread(
             target=run_analysis,
@@ -421,19 +457,22 @@ class ReviewDescriptionProcessor(PostProcessorApi):
                 camera_config,
                 final_data,
                 thumbs,
+                captions,
                 camera_config.review.genai,
                 sorted(self.config.all_labels),
                 self.config.all_attributes,
             ),
         ).start()
 
-    def save_debug_recording_frames(self, review_id: str, thumbs: list[bytes]) -> None:
+    def save_debug_recording_frames(
+        self, review_id: str, frames: list[tuple[bytes, float]]
+    ) -> None:
         """Write the recording frames sent to the provider out for debugging."""
         Path(os.path.join(CLIPS_DIR, "genai-requests", review_id)).mkdir(
             parents=True, exist_ok=True
         )
 
-        for idx, frame_bytes in enumerate(thumbs):
+        for idx, (frame_bytes, _) in enumerate(frames):
             with open(
                 os.path.join(CLIPS_DIR, f"genai-requests/{review_id}/{idx}.jpg"),
                 "wb",
@@ -445,7 +484,9 @@ class ReviewDescriptionProcessor(PostProcessorApi):
         camera: str,
         start_time: float,
         end_time: float,
-    ) -> list[str]:
+        frame_mode: ReviewFrameModeEnum = ReviewFrameModeEnum.frames,
+    ) -> list[tuple[str, float]]:
+        """Preview frame paths paired with the time each one was captured."""
         preview_dir = os.path.join(CACHE_DIR, "preview_frames")
         file_start = f"preview_{camera}-"
         start_file = f"{file_start}{start_time}.webp"
@@ -477,18 +518,29 @@ class ReviewDescriptionProcessor(PostProcessorApi):
 
         frame_count = len(all_frames)
         desired_frame_count = self.calculate_frame_count(
-            camera, duration=end_time - start_time
+            camera,
+            duration=end_time - start_time,
+            frame_mode=frame_mode,
         )
 
+        def with_timestamp(path: str) -> tuple[str, float]:
+            # Preview frames are named preview_<camera>-<timestamp>.webp
+            stem = os.path.basename(path).removesuffix(".webp")
+
+            try:
+                return (path, float(stem.removeprefix(file_start)))
+            except ValueError:
+                return (path, start_time)
+
         if frame_count <= desired_frame_count:
-            return all_frames
+            return [with_timestamp(f) for f in all_frames]
 
         selected_frames = []
         step_size = (frame_count - 1) / (desired_frame_count - 1)
 
         for i in range(desired_frame_count):
             index = round(i * step_size)
-            selected_frames.append(all_frames[index])
+            selected_frames.append(with_timestamp(all_frames[index]))
 
         return selected_frames
 
@@ -498,11 +550,12 @@ class ReviewDescriptionProcessor(PostProcessorApi):
         start_time: float,
         end_time: float,
         height: int = 480,
-    ) -> list[bytes]:
-        """Get frames from recordings at specified timestamps."""
+        frame_mode: ReviewFrameModeEnum = ReviewFrameModeEnum.frames,
+    ) -> list[tuple[bytes, float]]:
+        """Get frames from recordings paired with the time each was captured."""
         duration = end_time - start_time
         desired_frame_count = self.calculate_frame_count(
-            camera, duration, ImageSourceEnum.recordings, height
+            camera, duration, ImageSourceEnum.recordings, height, frame_mode
         )
 
         # Calculate evenly spaced timestamps throughout the duration
@@ -540,7 +593,7 @@ class ReviewDescriptionProcessor(PostProcessorApi):
             except DoesNotExist:
                 return None
 
-        frames = []
+        frames: list[tuple[bytes, float]] = []
 
         for timestamp in timestamps:
             try:
@@ -553,7 +606,7 @@ class ReviewDescriptionProcessor(PostProcessorApi):
                     image_data = extract_frame_from_recording(rounded_timestamp)
 
                 if image_data:
-                    frames.append(image_data)
+                    frames.append((image_data, timestamp))
                 else:
                     logger.warning(
                         f"No recording found for {camera} at timestamp {timestamp}"
@@ -574,7 +627,8 @@ class ReviewDescriptionProcessor(PostProcessorApi):
         thumb_path_fallback: str,
         review_id: str,
         save_debug: bool,
-    ) -> list[bytes]:
+        frame_mode: ReviewFrameModeEnum = ReviewFrameModeEnum.frames,
+    ) -> list[tuple[bytes, float]]:
         """Get preview frames and convert them to JPEG bytes.
 
         Args:
@@ -586,14 +640,14 @@ class ReviewDescriptionProcessor(PostProcessorApi):
             save_debug: Whether to save debug thumbnails
 
         Returns:
-            List of JPEG image bytes
+            List of (JPEG image bytes, capture timestamp) pairs
         """
-        frame_paths = self.get_cache_frames(camera, start_time, end_time)
+        frame_paths = self.get_cache_frames(camera, start_time, end_time, frame_mode)
         if not frame_paths:
-            frame_paths = [thumb_path_fallback]
+            frame_paths = [(thumb_path_fallback, start_time)]
 
-        thumbs = []
-        for idx, thumb_path in enumerate(frame_paths):
+        thumbs: list[tuple[bytes, float]] = []
+        for idx, (thumb_path, timestamp) in enumerate(frame_paths):
             thumb_data = cv2.imread(thumb_path)
 
             if thumb_data is None:
@@ -606,7 +660,7 @@ class ReviewDescriptionProcessor(PostProcessorApi):
                 ".jpg", thumb_data, [int(cv2.IMWRITE_JPEG_QUALITY), 100]
             )
             if ret:
-                thumbs.append(jpg.tobytes())
+                thumbs.append((jpg.tobytes(), timestamp))
 
             if save_debug:
                 Path(os.path.join(CLIPS_DIR, "genai-requests", review_id)).mkdir(
@@ -641,6 +695,7 @@ def run_analysis(
     camera_config: CameraConfig,
     final_data: dict[str, Any],
     thumbs: list[bytes],
+    frame_captions: list[str],
     genai_config: GenAIReviewConfig,
     labelmap_objects: list[str],
     attribute_labels: list[str],
@@ -697,6 +752,7 @@ def run_analysis(
         genai_config.debug_save_thumbnails,
         genai_config.activity_context_prompt,
         genai_config.response_style,
+        frame_captions,
     )
     review_inference_speed.update(datetime.datetime.now().timestamp() - start)
 

frigate/data_processing/real_time/audio_transcription.py

@@ -1,16 +1,19 @@
 """Handle processing audio for speech transcription using sherpa-onnx with FFmpeg pipe."""
 
+import collections
 import logging
 import os
 import queue
 import threading
-from typing import Any
+import time
+from typing import TYPE_CHECKING, Any
 
 import numpy as np
 
 from frigate.comms.inter_process import InterProcessRequestor
 from frigate.config import CameraConfig, FrigateConfig
-from frigate.const import MODEL_CACHE_DIR
+from frigate.config.classification import AudioTranscriptionModelEnum
+from frigate.const import AUDIO_DURATION, MODEL_CACHE_DIR
 from frigate.data_processing.common.audio_transcription.model import (
     AudioTranscriptionModelRunner,
 )
@@ -18,12 +21,38 @@ from frigate.data_processing.real_time.whisper_online import (
     FasterWhisperASR,
     OnlineASRProcessor,
 )
+from frigate.util.audio import (
+    clean_transcript,
+    pcm16_to_wav,
+    resolve_language,
+    stitch_transcripts,
+)
 
 from ..types import DataProcessorMetrics
 from .api import RealTimeProcessorApi
 
+if TYPE_CHECKING:
+    # importing frigate.genai eagerly would pull the provider SDKs into the
+    # audio process even when transcription runs on a local model
+    from frigate.genai.manager import GenAIClientManager
+
 logger = logging.getLogger(__name__)
 
+# Number of ~0.975s audio detector chunks per GenAI request. The window advances
+# one chunk at a time, so two chunks means a 50% overlap: every word lands whole
+# in at least one window, which whisper-family models need to avoid hallucinating
+# on a clipped clip. The cadence is fixed by the audio detector's frame size, so
+# this is a constant rather than a config knob.
+GENAI_WINDOW_CHUNKS = 2
+
+# Bound the queue at ~30s of audio so a slow or hung provider cannot grow it
+# without limit. The producer is the ffmpeg read thread and must never block.
+AUDIO_QUEUE_MAXSIZE = int(30 / AUDIO_DURATION)
+
+# A backed-up queue drops a chunk per cycle, so warning on each one would spam
+# the log once a second per camera for as long as the provider stays slow.
+AUDIO_DROP_WARN_INTERVAL = 10.0
+
 
 class AudioTranscriptionRealTimeProcessor(RealTimeProcessorApi):
     def __init__(
@@ -31,9 +60,10 @@ class AudioTranscriptionRealTimeProcessor(RealTimeProcessorApi):
         config: FrigateConfig,
         camera_config: CameraConfig,
         requestor: InterProcessRequestor,
-        model_runner: AudioTranscriptionModelRunner,
+        model_runner: AudioTranscriptionModelRunner | None,
         metrics: DataProcessorMetrics,
         stop_event: threading.Event,
+        genai_manager: "GenAIClientManager | None" = None,
     ):
         super().__init__(config, metrics)
         self.config = config
@@ -42,11 +72,31 @@ class AudioTranscriptionRealTimeProcessor(RealTimeProcessorApi):
         self.stream: Any = None
         self.whisper_model: FasterWhisperASR | None = None
         self.model_runner = model_runner
+        self.genai_manager = genai_manager
         self.transcription_segments: list[str] = []
-        self.audio_queue: queue.Queue[tuple[dict[str, Any], np.ndarray]] = queue.Queue()
+        self.audio_queue: queue.Queue[tuple[dict[str, Any], np.ndarray]] = queue.Queue(
+            maxsize=AUDIO_QUEUE_MAXSIZE
+        )
         self.stop_event = stop_event
+        self._use_genai = not isinstance(
+            config.audio_transcription.model, AudioTranscriptionModelEnum
+        )
+        # sliding window of raw int16 chunks; the deque's maxlen is what evicts
+        # the oldest chunk and so produces the overlap
+        self._genai_window: collections.deque[np.ndarray] = collections.deque(
+            maxlen=GENAI_WINDOW_CHUNKS
+        )
+        self._genai_committed = ""
+        # set by the producer when it discards a chunk, so the consumer knows the
+        # audio it is about to receive is not contiguous with what it buffered
+        self._audio_dropped = threading.Event()
+        self._last_drop_warning = 0.0
 
     def __build_recognizer(self) -> None:
+        if self._use_genai:
+            # nothing local to load; never import sherpa or FasterWhisperASR
+            return
+
         try:
             if self.config.audio_transcription.model_size == "large":
                 # Whisper models need to be per-process and can only run one stream at a time
@@ -64,7 +114,7 @@ class AudioTranscriptionRealTimeProcessor(RealTimeProcessorApi):
                 self.stream = OnlineASRProcessor(
                     asr=self.whisper_model,
                 )
-            else:
+            elif self.model_runner is not None:
                 logger.debug(f"Loading sherpa stream for {self.camera_config.name}")
                 self.stream = self.model_runner.model.create_stream()
             logger.debug(
@@ -76,6 +126,15 @@ class AudioTranscriptionRealTimeProcessor(RealTimeProcessorApi):
             )
 
     def __process_audio_stream(self, audio_data: np.ndarray) -> tuple[str, bool] | None:
+        # must precede both the model_runner guard (model_runner is None on this
+        # path) and the float32 normalization below (GenAI wants untouched int16)
+        if self._use_genai:
+            return self.__process_audio_genai(audio_data)
+
+        if self.model_runner is None:
+            logger.debug("Audio transcription (live) model runner not initialized")
+            return None
+
         if (
             self.model_runner.model is None
             and self.config.audio_transcription.model_size == "small"
@@ -140,6 +199,82 @@ class AudioTranscriptionRealTimeProcessor(RealTimeProcessorApi):
             logger.error(f"Error processing audio stream: {e}")
             return None
 
+    def __process_audio_genai(self, audio_data: np.ndarray) -> tuple[str, bool] | None:
+        """Transcribe a sliding overlapped window through the GenAI provider."""
+        client = self.genai_manager.transcribe_client if self.genai_manager else None
+
+        if not client:
+            logger.error(
+                "audio_transcription.model is '%s' (GenAI provider) but no transcribe "
+                "client is configured. Ensure the GenAI provider has 'transcribe' in its roles",
+                self.config.audio_transcription.model,
+            )
+            return None
+
+        if self._audio_dropped.is_set():
+            self._audio_dropped.clear()
+
+            # Chunks were discarded between what is buffered and this one, so
+            # concatenating them would splice non-adjacent audio into one window
+            # and destroy the overlap the stitcher depends on.
+            self._genai_window.clear()
+
+            if self._genai_committed:
+                # the transcript has a gap in it; close the utterance out rather
+                # than stitching across missing speech
+                return self.__end_genai_utterance()
+
+        self._genai_window.append(audio_data)
+
+        if len(self._genai_window) < GENAI_WINDOW_CHUNKS:
+            # wait for a full window so the first request is never a clipped clip
+            return None
+
+        window = np.concatenate(list(self._genai_window))
+
+        # Silence gate, using the same threshold audio detection uses. Gate the
+        # whole window rather than individual chunks; this is the primary cost
+        # and privacy brake and is what keeps a quiet camera near zero requests.
+        window_as_float = window.astype(np.float32)
+        rms = float(np.sqrt(np.mean(np.absolute(np.square(window_as_float)))))
+
+        if rms < self.camera_config.audio.min_volume:
+            logger.debug(
+                f"Window RMS {rms:.1f} below min_volume, skipping transcription"
+            )
+            return self.__end_genai_utterance()
+
+        text = client.transcribe(
+            pcm16_to_wav(window),
+            language=resolve_language(self.config.audio_transcription.language),
+        )
+
+        # cleaning has to come first: a silent window often comes back as the
+        # model's preamble alone, which is silence, not a word to commit
+        cleaned = clean_transcript(text)
+
+        if not cleaned:
+            return self.__end_genai_utterance()
+
+        self._genai_committed = stitch_transcripts(self._genai_committed, cleaned)
+
+        # no VAD on this path, so mirror the whisper branch's heuristic endpoint
+        is_endpoint = (
+            self._genai_committed.endswith((".", "!", "?"))
+            and len(self._genai_committed) > 300
+        )
+
+        logger.debug(f"GenAI transcription: '{self._genai_committed}'")
+
+        return self._genai_committed, is_endpoint
+
+    def __end_genai_utterance(self) -> tuple[str, bool] | None:
+        """Close out the current utterance when a window carries no speech."""
+        if not self._genai_committed:
+            return None
+
+        return self._genai_committed, True
+
     def process_frame(self, obj_data: dict[str, Any], frame: np.ndarray) -> None:
         pass
 
@@ -148,8 +283,38 @@ class AudioTranscriptionRealTimeProcessor(RealTimeProcessorApi):
             logger.debug("No audio data provided for transcription")
             return None
 
-        # enqueue audio data for processing in the thread
-        self.audio_queue.put((obj_data, audio))
+        # enqueue audio data for processing in the thread. never block: the
+        # producer is the ffmpeg read thread that audio detection depends on,
+        # so on a backlog drop the oldest chunk instead.
+        try:
+            self.audio_queue.put_nowait((obj_data, audio))
+        except queue.Full:
+            try:
+                self.audio_queue.get_nowait()
+                self.audio_queue.task_done()
+            except queue.Empty:
+                pass
+
+            # the stream now has a hole in it, which the consumer has to know
+            # about before it splices the next chunk onto what it already holds
+            self._audio_dropped.set()
+
+            now = time.monotonic()
+
+            if now - self._last_drop_warning >= AUDIO_DROP_WARN_INTERVAL:
+                self._last_drop_warning = now
+                logger.warning(
+                    "Audio transcription queue for %s is full, dropping audio. The "
+                    "provider is not keeping up with the %.2fs chunk rate",
+                    self.camera_config.name,
+                    AUDIO_DURATION,
+                )
+
+            try:
+                self.audio_queue.put_nowait((obj_data, audio))
+            except queue.Full:
+                pass
+
         return None
 
     def run(self) -> None:
@@ -205,6 +370,14 @@ class AudioTranscriptionRealTimeProcessor(RealTimeProcessorApi):
                 break
 
     def reset(self) -> None:
+        if self._use_genai:
+            self._genai_committed = ""
+            # stale audio carried across an utterance boundary would be
+            # re-transcribed into the next one
+            self._genai_window.clear()
+            logger.debug("Stream reset")
+            return
+
         if self.config.audio_transcription.model_size == "large":
             # get final output from whisper
             output = self.stream.finish()
@@ -218,7 +391,7 @@ class AudioTranscriptionRealTimeProcessor(RealTimeProcessorApi):
             # reset whisper
             self.stream.init()
             self.transcription_segments = []
-        else:
+        elif self.model_runner is not None:
             # reset sherpa
             self.model_runner.model.reset(self.stream)
 
@@ -226,6 +399,24 @@ class AudioTranscriptionRealTimeProcessor(RealTimeProcessorApi):
 
     def check_unload_model(self) -> None:
         # regularly called in the loop in audio maintainer
+        if self._use_genai:
+            # no model to unload, but this is the hook that fires when
+            # live_enabled flips off. guard on emptiness: called ~1x/s per camera.
+            if self._genai_committed or self._genai_window:
+                logger.debug(
+                    f"Clearing GenAI transcription state for {self.camera_config.name}"
+                )
+                self.clear_audio_queue()
+                self._genai_committed = ""
+                self._genai_window.clear()
+
+                self.requestor.send_data(
+                    f"{self.camera_config.name}/audio/transcription",
+                    "",
+                )
+
+            return
+
         if (
             self.config.audio_transcription.model_size == "large"
             and self.whisper_model is not None
@@ -270,6 +461,10 @@ class AudioTranscriptionRealTimeProcessor(RealTimeProcessorApi):
         self, topic: str, request_data: dict[str, Any]
     ) -> dict[str, Any] | None:
         if topic == "clear_audio_recognizer":
+            if self._use_genai:
+                self.reset()
+                return {"message": "Audio transcription state cleared", "success": True}
+
             self.stream = None
             self.__build_recognizer()
             return {"message": "Audio recognizer cleared and rebuilt", "success": True}

frigate/detectors/hardware.py

@@ -273,6 +273,16 @@ def detect_memryx() -> DetectionHardware | None:
     return _hardware("memryx", "memryx", "MemryX MX3", units)
 
 
+def detect_deepx() -> DetectionHardware | None:
+    """Find DEEPX NPUs by their device nodes."""
+    units = _dev_units("dxrt*", "deepx:PCIe:{index}", "PCIe")
+
+    if not units:
+        return None
+
+    return _hardware("deepx", "deepx", "DEEPX NPU", units)
+
+
 def detect_rockchip() -> DetectionHardware | None:
     """Find a Rockchip NPU by reading the SoC from the device tree."""
     compatible = _read(f"{PROC_ROOT}/device-tree/compatible")
@@ -319,6 +329,7 @@ PROBES = (
     detect_coral_usb,
     detect_hailo,
     detect_memryx,
+    detect_deepx,
     detect_intel_npu,
     detect_intel_gpu,
     detect_nvidia_gpu,

frigate/embeddings/maintainer.py

@@ -68,7 +68,7 @@ from frigate.events.types import (
     RegenerateDescriptionEnum,
 )
 from frigate.genai import GenAIClientManager
-from frigate.models import Event, Recordings, ReviewSegment, Trigger
+from frigate.models import Event, Recordings, ReviewSegment, Timeline, Trigger
 from frigate.types import TrackedObjectUpdateTypesEnum
 from frigate.util.builtin import serialize
 from frigate.util.file import get_event_thumbnail_bytes
@@ -139,7 +139,7 @@ class EmbeddingMaintainer(threading.Thread):
             ),
             load_vec_extension=True,
         )
-        models = [Event, Recordings, ReviewSegment, Trigger]
+        models = [Event, Recordings, ReviewSegment, Timeline, Trigger]
         db.bind(models)
 
         self.genai_manager = GenAIClientManager(config)
@@ -251,7 +251,11 @@ class EmbeddingMaintainer(threading.Thread):
         ):
             self.post_processors.append(
                 AudioTranscriptionPostProcessor(
-                    self.config, self.requestor, self.embeddings, metrics
+                    self.config,
+                    self.requestor,
+                    self.embeddings,
+                    metrics,
+                    self.genai_manager,
                 )
             )
 

frigate/events/audio.py

@@ -11,6 +11,7 @@ from typing import Any
 
 import numpy as np
 
+from frigate.camera import CameraMetrics
 from frigate.comms.detections_updater import DetectionPublisher, DetectionTypeEnum
 from frigate.comms.inter_process import InterProcessRequestor
 from frigate.config import CameraConfig, CameraInput, FrigateConfig
@@ -19,6 +20,7 @@ from frigate.config.camera.updater import (
     CameraConfigUpdateEnum,
     CameraConfigUpdateSubscriber,
 )
+from frigate.config.classification import AudioTranscriptionModelEnum
 from frigate.const import (
     AUDIO_DURATION,
     AUDIO_FORMAT,
@@ -35,6 +37,7 @@ from frigate.data_processing.common.audio_transcription.model import (
 from frigate.data_processing.real_time.audio_transcription import (
     AudioTranscriptionRealTimeProcessor,
 )
+from frigate.data_processing.types import DataProcessorMetrics
 from frigate.ffmpeg_presets import parse_preset_input
 from frigate.log import LogPipe, suppress_stderr_during
 from frigate.util.builtin import get_ffmpeg_arg_list, load_labels
@@ -85,6 +88,7 @@ class AudioProcessor(FrigateProcess):
         self,
         config: FrigateConfig,
         camera_metrics: DictProxy,
+        embeddings_metrics: DataProcessorMetrics,
         stop_event: MpEvent,
     ):
         super().__init__(
@@ -92,8 +96,38 @@ class AudioProcessor(FrigateProcess):
         )
 
         self.camera_metrics = camera_metrics
+        self.embeddings_metrics = embeddings_metrics
         self.config = config
 
+    def spawn_if_needed(self, camera: CameraConfig) -> None:
+        """Start an audio maintainer for the camera once everything it needs
+        has arrived. Returning early leaves the camera for the next poll."""
+        name = camera.name
+        if name is None or name in self.audio_threads:
+            return
+        if not camera.enabled or not camera.audio.enabled:
+            return
+        # ffmpeg update may not have arrived yet
+        if not any("audio" in i.roles for i in camera.ffmpeg.inputs):
+            return
+        # the camera maintainer creates metrics on its own poll of the same
+        # add update and may not have gotten there yet
+        metrics = self.camera_metrics.get(name)
+        if metrics is None:
+            return
+        thread = AudioEventMaintainer(
+            camera,
+            self.config,
+            metrics,
+            self.embeddings_metrics,
+            self.transcription_model_runner,
+            self.stop_event,  # type: ignore[arg-type]
+            self.genai_manager,
+        )
+        self.audio_threads[name] = thread
+        thread.start()
+        self.logger.info(f"Audio maintainer started for {name}")
+
     def __stop_audio_thread(self, camera: str) -> None:
         thread = self.audio_threads.pop(camera, None)
         if thread is None:
@@ -112,18 +146,31 @@ class AudioProcessor(FrigateProcess):
 
         threading.current_thread().name = "process:audio_manager"
 
+        self.transcription_model_runner: AudioTranscriptionModelRunner | None = None
+        self.genai_manager: Any = None
+
         if any(
             c.enabled_in_config and c.audio_transcription.enabled
             for c in self.config.cameras.values()
         ):
-            self.transcription_model_runner: AudioTranscriptionModelRunner | None = (
-                AudioTranscriptionModelRunner(
+            if isinstance(
+                self.config.audio_transcription.model, AudioTranscriptionModelEnum
+            ):
+                # AudioTranscriptionModelRunner.__init__ unconditionally fetches
+                # sherpa-onnx or whisper weights, so only build it on the local path
+                self.transcription_model_runner = AudioTranscriptionModelRunner(
                     self.config.audio_transcription.device or "AUTO",
                     self.config.audio_transcription.model_size,
                 )
-            )
-        else:
-            self.transcription_model_runner = None
+            else:
+                # imported here rather than at module scope: frigate.genai pulls in
+                # numpy, the provider SDKs, frigate.models, and the prompt builders,
+                # and this process runs at PROCESS_PRIORITY_HIGH. built after the
+                # fork because SDK clients hold sockets and TLS state that must not
+                # cross it; clients themselves stay lazy behind the role property.
+                from frigate.genai.manager import GenAIClientManager
+
+                self.genai_manager = GenAIClientManager(self.config)
 
         config_subscriber = CameraConfigUpdateSubscriber(
             self.config,
@@ -136,28 +183,8 @@ class AudioProcessor(FrigateProcess):
             ],
         )
 
-        def spawn_if_needed(camera: CameraConfig) -> None:
-            name = camera.name
-            if name is None or name in self.audio_threads:
-                return
-            if not camera.enabled or not camera.audio.enabled:
-                return
-            # ffmpeg update may not have arrived yet; wait for next poll
-            if not any("audio" in i.roles for i in camera.ffmpeg.inputs):
-                return
-            thread = AudioEventMaintainer(
-                camera,
-                self.config,
-                self.camera_metrics,
-                self.transcription_model_runner,
-                self.stop_event,  # type: ignore[arg-type]
-            )
-            self.audio_threads[name] = thread
-            thread.start()
-            self.logger.info(f"Audio maintainer started for {name}")
-
         for camera in self.config.cameras.values():
-            spawn_if_needed(camera)
+            self.spawn_if_needed(camera)
 
         self.logger.info(f"Audio processor started (pid: {self.pid})")
 
@@ -167,15 +194,14 @@ class AudioProcessor(FrigateProcess):
             updated_topics = config_subscriber.check_for_updates()
 
             # stop maintainers for removed cameras so their ffmpeg process is
-            # torn down and they stop touching camera_metrics (which the camera
-            # maintainer has already popped for the removed camera)
+            # torn down
             for removed_camera in updated_topics.get(
                 CameraConfigUpdateEnum.remove.name, []
             ):
                 self.__stop_audio_thread(removed_camera)
 
             for camera in self.config.cameras.values():
-                spawn_if_needed(camera)
+                self.spawn_if_needed(camera)
 
         config_subscriber.stop()
 
@@ -197,15 +223,21 @@ class AudioEventMaintainer(threading.Thread):
         self,
         camera: CameraConfig,
         config: FrigateConfig,
-        camera_metrics: DictProxy,
+        metrics: CameraMetrics,
+        embeddings_metrics: DataProcessorMetrics,
         audio_transcription_model_runner: AudioTranscriptionModelRunner | None,
         stop_event: threading.Event,
+        genai_manager: Any = None,
     ) -> None:
         super().__init__(name=f"{camera.name}_audio_event_processor")
 
         self.config = config
         self.camera_config = camera
-        self.camera_metrics = camera_metrics
+        # hold the metrics object rather than indexing the manager dict per
+        # chunk, which costs an IPC round trip and breaks once the camera
+        # maintainer pops the entry on removal
+        self.metrics = metrics
+        self.embeddings_metrics = embeddings_metrics
         self.stop_event = stop_event
         # per-camera stop signal so a single maintainer can be torn down at
         # runtime (e.g. on camera removal) without stopping the whole process
@@ -222,6 +254,7 @@ class AudioEventMaintainer(threading.Thread):
         self.logpipe = LogPipe(f"ffmpeg.{self.camera_config.name}.audio")
         self.audio_listener: subprocess.Popen[Any] | None = None
         self.audio_transcription_model_runner = audio_transcription_model_runner
+        self.genai_manager = genai_manager
         self.transcription_processor = None
         self.transcription_thread = None
 
@@ -238,9 +271,9 @@ class AudioEventMaintainer(threading.Thread):
         )
         self.detection_publisher = DetectionPublisher(DetectionTypeEnum.audio.value)
 
-        if (
-            self.camera_config.audio_transcription.enabled
-            and self.audio_transcription_model_runner is not None
+        if self.camera_config.audio_transcription.enabled and (
+            self.audio_transcription_model_runner is not None
+            or self.genai_manager is not None
         ):
             # init the transcription processor for this camera
             self.transcription_processor = AudioTranscriptionRealTimeProcessor(
@@ -248,8 +281,9 @@ class AudioEventMaintainer(threading.Thread):
                 camera_config=self.camera_config,
                 requestor=self.requestor,
                 model_runner=self.audio_transcription_model_runner,
-                metrics=self.camera_metrics[self.camera_config.name],
+                metrics=self.embeddings_metrics,
                 stop_event=self.stop_event,
+                genai_manager=self.genai_manager,
             )
 
             self.transcription_thread = threading.Thread(
@@ -273,8 +307,8 @@ class AudioEventMaintainer(threading.Thread):
         audio_as_float: np.ndarray = audio.astype(np.float32)
         rms, dBFS = self.calculate_audio_levels(audio_as_float)
 
-        self.camera_metrics[self.camera_config.name].audio_rms.value = rms
-        self.camera_metrics[self.camera_config.name].audio_dBFS.value = dBFS
+        self.metrics.audio_rms.value = rms
+        self.metrics.audio_dBFS.value = dBFS
 
         audio_detections: list[tuple[str, float]] = []
 
@@ -359,7 +393,6 @@ class AudioEventMaintainer(threading.Thread):
                 return
 
             time.sleep(self.camera_config.ffmpeg.retry_interval)
-            self.logpipe.dump()
             self.start_or_restart_ffmpeg()
 
         if self.audio_listener is None or self.audio_listener.stdout is None:

frigate/genai/__init__.py

@@ -105,8 +105,21 @@ class GenAIClient:
         debug_save: bool,
         activity_context_prompt: str,
         response_style: str = "default",
+        frame_captions: list[str] | None = None,
     ) -> ReviewMetadata | None:
-        """Generate a description for the review item activity."""
+        """Generate a description for the review item activity.
+
+        `frame_captions` holds one caption per thumbnail for the annotated
+        frame mode; each is sent directly before its frame.
+        """
+        if frame_captions and len(frame_captions) != len(thumbnails):
+            logger.warning(
+                "Got %d frame captions for %d thumbnails, sending plain frames",
+                len(frame_captions),
+                len(thumbnails),
+            )
+            frame_captions = None
+
         context_prompt = build_review_description_prompt(
             review_data,
             thumbnails,
@@ -114,6 +127,7 @@ class GenAIClient:
             preferred_language,
             activity_context_prompt,
             response_style,
+            frame_captions,
         )
 
         logger.debug(
@@ -129,9 +143,30 @@ class GenAIClient:
             ) as f:
                 f.write(context_prompt)
 
+            if frame_captions:
+                # One file per frame, numbered to match the image it precedes
+                # (0.txt goes with 0.jpg), so the debug folder replays without
+                # having to re-derive the mapping.
+                for index, caption in enumerate(frame_captions):
+                    with open(
+                        os.path.join(
+                            CLIPS_DIR,
+                            "genai-requests",
+                            review_data["id"],
+                            f"{index}.txt",
+                        ),
+                        "w",
+                    ) as f:
+                        f.write(caption)
+
         response_format = build_review_description_response_format(concerns)
 
-        response = self._send(context_prompt, thumbnails, response_format)
+        response = self._send(
+            context_prompt,
+            thumbnails,
+            response_format,
+            image_captions=frame_captions,
+        )
 
         if debug_save and response:
             with open(
@@ -269,6 +304,7 @@ class GenAIClient:
         images: list[bytes],
         response_format: dict | None = None,
         enable_thinking: bool = False,
+        image_captions: list[str] | None = None,
     ) -> str | None:
         """Submit a request to the provider.
 
@@ -276,6 +312,10 @@ class GenAIClient:
         ``supports_toggleable_thinking``. Description-style callers leave it
         at the default (off) since synthesis tasks don't benefit from
         reasoning traces.
+
+        ``image_captions`` carries one caption per image, to be placed
+        immediately before its image so the model can tell the frames apart.
+        Providers build their request order with ``interleave_images``.
         """
         return None
 
@@ -298,6 +338,11 @@ class GenAIClient:
         """Whether the configured model can generate embeddings via embed()."""
         return False
 
+    @property
+    def supports_transcription(self) -> bool:
+        """Whether the configured model can transcribe audio via transcribe()."""
+        return False
+
     def list_models(self) -> list[str]:
         """Return the list of model names available from this provider.
 
@@ -305,6 +350,21 @@ class GenAIClient:
         """
         return []
 
+    def list_model_capabilities(self) -> dict[str, dict[str, bool]]:
+        """Return capability flags for each model the provider serves.
+
+        Only providers whose backend advertises capabilities per model can
+        populate this; llama.cpp reports input modalities for every model it
+        serves, so one request describes them all. An empty mapping means "no
+        per-model information available", and callers fall back to this
+        client's own capability properties, which describe only the configured
+        model. A model absent from a non-empty mapping means the same thing.
+
+        Returns:
+            Model name (including aliases) to its capability flags
+        """
+        return {}
+
     def get_context_size(self) -> int:
         """Get the context window size for this provider in tokens."""
         return 4096
@@ -336,6 +396,33 @@ class GenAIClient:
         )
         return []
 
+    def transcribe(
+        self,
+        audio: bytes,
+        language: str | None = None,
+        mime_type: str = "audio/wav",
+    ) -> str | None:
+        """Transcribe speech audio to text.
+
+        Audio is passed as a self-describing blob rather than raw samples so
+        every provider receives a container it can declare, and WAV framing
+        lives in one place instead of in each plugin.
+
+        Args:
+            audio: The encoded audio payload (WAV bytes by default)
+            language: Optional ISO language hint for the provider
+            mime_type: Media type of ``audio``
+
+        Returns:
+            The transcript, or None when the provider cannot produce one
+        """
+        logger.warning(
+            "%s does not support transcription. "
+            "This method should be overridden by the provider implementation.",
+            self.__class__.__name__,
+        )
+        return None
+
     def chat_with_tools(
         self,
         messages: list[dict[str, Any]],

frigate/genai/manager.py

@@ -110,6 +110,12 @@ class GenAIClientManager:
         name = self._role_map.get(GenAIRoleEnum.embeddings)
         return self._get_client(name) if name else None
 
+    @property
+    def transcribe_client(self) -> "GenAIClient | None":
+        """Client configured for the transcribe role."""
+        name = self._role_map.get(GenAIRoleEnum.transcribe)
+        return self._get_client(name) if name else None
+
     def role_info(self) -> dict[str, dict[str, Any]]:
         """Return the model selected for each configured role and its context size.
 
@@ -144,5 +150,11 @@ class GenAIClientManager:
                 "roles": [r.value for r in genai_cfg.roles],
                 "supports_toggleable_thinking": client.supports_toggleable_thinking,
                 "supports_embeddings": client.supports_embeddings,
+                "supports_transcription": client.supports_transcription,
+                # Capabilities of the configured model are above; this maps every
+                # model the provider serves to its own, so the UI can react to a
+                # model selected but not yet saved. Empty when the provider
+                # cannot report capabilities without loading a model.
+                "model_capabilities": client.list_model_capabilities(),
             }
         return result

frigate/genai/plugins/azure-openai.py

@@ -13,6 +13,19 @@ overrides what is genuinely Azure-specific:
 - Context size: Azure does not expose a per-model ``max_model_len`` field
   reliably, so we keep the historical 128K default rather than the
   model-name heuristic used by OpenAI.
+
+Transcription is inherited too: :class:`openai.AzureOpenAI` exposes the same
+``audio.transcriptions.create``. Two Azure-specific caveats apply when using
+the ``transcribe`` role:
+
+- ``model`` must be the Azure *deployment* name, not the underlying model name.
+- The ``api-version`` parsed from ``base_url`` must be 2024-06-01 or later;
+  earlier versions have no transcriptions route and the 404 surfaces only as a
+  generic provider error.
+- Because ``model`` is a deployment name, the inherited check that picks
+  ``languages`` over ``language`` for gpt-transcribe cannot fire unless the
+  deployment happens to be named after the model. Name the deployment
+  ``gpt-transcribe`` to get the right field, or leave the language on ``auto``.
 """
 
 import logging

frigate/genai/plugins/gemini.py

@@ -13,9 +13,14 @@ from google.genai.types import FunctionCallingConfigMode
 
 from frigate.config import GenAIProviderEnum
 from frigate.genai import GenAIClient, register_genai_provider
+from frigate.genai.utils import interleave_images
 
 logger = logging.getLogger(__name__)
 
+# Gemini requests carrying inline data are capped at ~20 MB total; stay well
+# under it so the request fails as a log line rather than a 400.
+GEMINI_MAX_INLINE_BYTES = 15 * 1024 * 1024
+
 
 def _decode_thought_signature(value: Any) -> bytes | None:
     """Decode a base64-encoded thought_signature carried across conversation turns."""
@@ -118,11 +123,16 @@ class GeminiClient(GenAIClient):
         images: list[bytes],
         response_format: dict | None = None,
         enable_thinking: bool = False,
+        image_captions: list[str] | None = None,
     ) -> str | None:
         """Submit a request to Gemini."""
-        contents = [prompt] + [
-            types.Part.from_bytes(data=img, mime_type="image/jpeg") for img in images
+        contents: list[Any] = [
+            part
+            if isinstance(part, str)
+            else types.Part.from_bytes(data=part, mime_type="image/jpeg")
+            for part in interleave_images(prompt, images, image_captions)
         ]
+
         try:
             # Merge runtime_options into generation_config if provided
             generation_config_dict: dict[str, Any] = {"candidate_count": 1}
@@ -136,7 +146,7 @@ class GeminiClient(GenAIClient):
 
             response = self.provider.models.generate_content(
                 model=self.genai_config.model,
-                contents=contents,  # type: ignore[arg-type]
+                contents=contents,
                 config=types.GenerateContentConfig(
                     **generation_config_dict,
                 ),
@@ -157,6 +167,58 @@ class GeminiClient(GenAIClient):
             return None
         return description
 
+    @property
+    def supports_transcription(self) -> bool:
+        """Gemini models accept inline audio parts."""
+        return True
+
+    def transcribe(
+        self,
+        audio: bytes,
+        language: str | None = None,
+        mime_type: str = "audio/wav",
+    ) -> str | None:
+        """Transcribe audio by sending it as an inline part alongside a prompt."""
+        if len(audio) > GEMINI_MAX_INLINE_BYTES:
+            logger.warning(
+                "Audio payload of %d bytes exceeds the Gemini inline limit; skipping transcription",
+                len(audio),
+            )
+            return None
+
+        prompt = "Transcribe the speech in this audio verbatim. Respond with the transcript only, and with nothing at all if there is no speech."
+
+        if language:
+            prompt += f" The speech is in language '{language}'."
+
+        try:
+            contents: list[Any] = [
+                prompt,
+                types.Part.from_bytes(data=audio, mime_type=mime_type),
+            ]
+            response = self.provider.models.generate_content(
+                model=self.genai_config.model,
+                contents=contents,
+                config=types.GenerateContentConfig(candidate_count=1),
+            )
+        except errors.APIError as e:
+            logger.warning("Gemini returned an error: %s", str(e))
+            return None
+        except Exception as e:
+            logger.warning("An unexpected error occurred with Gemini: %s", str(e))
+            return None
+
+        try:
+            if response.text is None:
+                return None
+
+            transcript = response.text.strip()
+        except (ValueError, AttributeError):
+            # No transcript was generated
+            return None
+
+        return transcript or None
+
     def list_models(self) -> list[str]:
         """Return available model names from Gemini."""
         try:

frigate/genai/plugins/llama_cpp.py

@@ -14,7 +14,7 @@ from PIL import Image
 
 from frigate.config import GenAIProviderEnum
 from frigate.genai import GenAIClient, register_genai_provider
-from frigate.genai.utils import parse_tool_calls_from_message
+from frigate.genai.utils import interleave_images, parse_tool_calls_from_message
 
 logger = logging.getLogger(__name__)
 
@@ -333,6 +333,7 @@ class LlamaCppClient(GenAIClient):
         images: list[bytes],
         response_format: dict | None = None,
         enable_thinking: bool = False,
+        image_captions: list[str] | None = None,
     ) -> str | None:
         """Submit a request to llama.cpp server."""
         if self.provider is None:
@@ -342,18 +343,17 @@ class LlamaCppClient(GenAIClient):
             return None
 
         try:
-            content = [
-                {
-                    "type": "text",
-                    "text": prompt,
-                }
-            ]
-            for image in images:
-                encoded_image = base64.b64encode(image).decode("utf-8")
+            content: list[dict[str, Any]] = []
+            for part in interleave_images(prompt, images, image_captions):
+                if isinstance(part, str):
+                    content.append({"type": "text", "text": part})
+                    continue
+
+                encoded_image = base64.b64encode(part).decode("utf-8")
                 content.append(
                     {
                         "type": "image_url",
-                        "image_url": {  # type: ignore[dict-item]
+                        "image_url": {
                             "url": f"data:image/jpeg;base64,{encoded_image}",
                         },
                     }
@@ -408,6 +408,126 @@ class LlamaCppClient(GenAIClient):
         """Whether the loaded model supports audio input."""
         return self._supports_audio
 
+    @property
+    def supports_transcription(self) -> bool:
+        """Audio-capable models can transcribe through chat completions."""
+        return self._supports_audio
+
+    def transcribe(
+        self,
+        audio: bytes,
+        language: str | None = None,
+        mime_type: str = "audio/wav",
+    ) -> str | None:
+        """Transcribe audio through the OpenAI-compatible transcriptions route.
+
+        llama.cpp serves /v1/audio/transcriptions for any audio-capable model,
+        not only a separately loaded whisper (ggml-org/llama.cpp#21863), so it
+        covers exactly the models supports_transcription detects. It takes the
+        language as a native multipart field, which is the only thing dedicated
+        ASR models honor: they read the chat prompt as contextual biasing, so
+        asking one there to use a language does nothing.
+
+        Falls back to chat completions when the server predates that route.
+        """
+        if self.provider is None:
+            logger.warning(
+                "llama.cpp provider has not been initialized, audio will not be transcribed. Check your llama.cpp configuration."
+            )
+            return None
+
+        if not self._supports_audio:
+            logger.warning(
+                "llama.cpp model '%s' does not accept audio input",
+                self.genai_config.model,
+            )
+            return None
+
+        try:
+            data = {"model": self.genai_config.model, "response_format": "json"}
+
+            if language:
+                data["language"] = language
+
+            response = self._post(
+                f"{self.provider}/v1/audio/transcriptions",
+                files={"file": ("audio.wav", audio, mime_type)},
+                data=data,
+                timeout=self.timeout,
+            )
+
+            if response.status_code == 404:
+                logger.debug(
+                    "llama.cpp server has no /v1/audio/transcriptions route, using chat completions"
+                )
+                return self._transcribe_via_chat(audio, language)
+
+            response.raise_for_status()
+            result = response.json()
+            text = result.get("text") if isinstance(result, dict) else None
+
+            return str(text).strip() or None if text else None
+        except Exception as e:
+            logger.warning("llama.cpp returned an error: %s", str(e))
+            return None
+
+    def _transcribe_via_chat(self, audio: bytes, language: str | None) -> str | None:
+        """Transcribe through /v1/chat/completions, for servers without the
+        transcriptions route.
+
+        The _media_marker / multimodal_data convention is an /embeddings-only
+        protocol, so no marker-refresh retry is needed here.
+        """
+        prompt = "Transcribe the speech in this audio verbatim. Respond with the transcript only, and with nothing at all if there is no speech."
+
+        if language:
+            prompt += f" The speech is in language '{language}'."
+
+        try:
+            encoded_audio = base64.b64encode(audio).decode("utf-8")
+            payload: dict[str, Any] = {
+                "model": self.genai_config.model,
+                "messages": [
+                    {
+                        "role": "user",
+                        "content": [
+                            {"type": "text", "text": prompt},
+                            {
+                                "type": "input_audio",
+                                "input_audio": {
+                                    "data": encoded_audio,
+                                    "format": "wav",
+                                },
+                            },
+                        ],
+                    },
+                ],
+                **self.provider_options,
+            }
+
+            response = self._post(
+                f"{self.provider}/v1/chat/completions",
+                json=payload,
+                timeout=self.timeout,
+            )
+            response.raise_for_status()
+            result = response.json()
+
+            if (
+                result is not None
+                and "choices" in result
+                and len(result["choices"]) > 0
+            ):
+                choice = result["choices"][0]
+
+                if "message" in choice and choice["message"].get("content"):
+                    return str(choice["message"]["content"].strip()) or None
+
+            return None
+        except Exception as e:
+            logger.warning("llama.cpp returned an error: %s", str(e))
+            return None
+
     @property
     def supports_tools(self) -> bool:
         """Whether the loaded model supports tool/function calling."""
@@ -417,28 +537,73 @@ class LlamaCppClient(GenAIClient):
     def supports_toggleable_thinking(self) -> bool:
         return self._supports_reasoning
 
-    def list_models(self) -> list[str]:
-        """Return available model IDs from the llama.cpp server."""
+    def _fetch_models_data(self) -> list[dict[str, Any]]:
+        """Return the raw /v1/models entries, or an empty list if unreachable."""
         base_url = self.provider or (
             self.genai_config.base_url.rstrip("/")
             if self.genai_config.base_url
             else None
         )
+
         if base_url is None:
             return []
+
         try:
             response = self._get(f"{base_url}/v1/models", timeout=10)
             response.raise_for_status()
-            models = []
-            for m in response.json().get("data", []):
-                models.append(m.get("id", "unknown"))
-                for alias in m.get("aliases", []):
-                    models.append(alias)
-            return sorted(models)
+            data = response.json().get("data", [])
         except Exception as e:
             logger.warning("Failed to list llama.cpp models: %s", e)
             return []
 
+        return data if isinstance(data, list) else []
+
+    def list_models(self) -> list[str]:
+        """Return available model IDs from the llama.cpp server."""
+        models: set[str] = set()
+
+        # llama-server lists the id among the aliases when --alias is set
+        for m in self._fetch_models_data():
+            models.add(m.get("id", "unknown"))
+            models.update(m.get("aliases", []))
+
+        return sorted(models)
+
+    def list_model_capabilities(self) -> dict[str, dict[str, bool]]:
+        """Report input modalities for every model the server serves.
+
+        Since ggml-org/llama.cpp#22952 each /v1/models entry carries
+        architecture.input_modalities, so a single request describes every
+        model rather than just the configured one. That is what lets the UI
+        answer "can the model I just picked transcribe" before the config is
+        saved and a client for it exists.
+
+        Models whose entry predates that field are omitted rather than reported
+        as incapable, so an older server falls back to the /props probe instead
+        of silently losing capabilities it actually has.
+        """
+        capabilities: dict[str, dict[str, bool]] = {}
+
+        for model in self._fetch_models_data():
+            architecture = model.get("architecture") or {}
+            modalities = architecture.get("input_modalities")
+
+            if not isinstance(modalities, list) or not modalities:
+                continue
+
+            flags = {
+                "supports_vision": "image" in modalities,
+                "supports_transcription": "audio" in modalities,
+            }
+
+            names = [model.get("id"), *(model.get("aliases") or [])]
+
+            for name in names:
+                if isinstance(name, str) and name:
+                    capabilities[name] = flags
+
+        return capabilities
+
     def get_context_size(self) -> int:
         """Get the context window size for llama.cpp.
 

frigate/genai/plugins/ollama.py

@@ -14,7 +14,7 @@ from ollama import ResponseError
 
 from frigate.config import GenAIProviderEnum
 from frigate.genai import GenAIClient, register_genai_provider
-from frigate.genai.utils import parse_tool_calls_from_message
+from frigate.genai.utils import interleave_images, parse_tool_calls_from_message
 
 logger = logging.getLogger(__name__)
 
@@ -50,6 +50,28 @@ def _extract_ollama_stats(response: Any) -> dict[str, Any] | None:
     return stats or None
 
 
+# Ollama replaces each occurrence of this marker in a message, in order, with
+# the next image from the message's images list. Without markers it puts every
+# image before the text.
+IMAGE_PLACEHOLDER = "[img]"
+
+
+def _flatten_parts(parts: list[str | bytes]) -> tuple[str, list[bytes] | None]:
+    """Collapse ordered text and image parts into Ollama's (content, images)
+    shape, marking where each image goes so the order survives."""
+    text: list[str] = []
+    images: list[bytes] = []
+
+    for part in parts:
+        if isinstance(part, bytes):
+            text.append(IMAGE_PLACEHOLDER)
+            images.append(part)
+        elif part:
+            text.append(part)
+
+    return "\n".join(text), (images or None)
+
+
 def _normalize_multimodal_content(
     content: Any,
 ) -> tuple[str | None, list[bytes] | None]:
@@ -58,13 +80,13 @@ def _normalize_multimodal_content(
     The chat API constructs user messages with content as a list of
     ``{"type": "text"}`` and ``{"type": "image_url"}`` parts when a tool
     returns a live frame. Ollama's SDK requires content to be a string and
-    images to be passed in a separate field, so we extract each.
+    images to be passed in a separate field, so images are pulled out and
+    their positions marked with placeholders.
     """
     if not isinstance(content, list):
         return content, None
 
-    text_parts: list[str] = []
-    images: list[bytes] = []
+    parts: list[str | bytes] = []
     for part in content:
         if not isinstance(part, dict):
             continue
@@ -72,17 +94,20 @@ def _normalize_multimodal_content(
         if part_type == "text":
             text = part.get("text")
             if text:
-                text_parts.append(str(text))
+                parts.append(str(text))
         elif part_type == "image_url":
             url = (part.get("image_url") or {}).get("url", "")
             if isinstance(url, str) and url.startswith("data:"):
                 try:
                     encoded = url.split(",", 1)[1]
-                    images.append(base64.b64decode(encoded, validate=True))
+                    parts.append(base64.b64decode(encoded, validate=True))
                 except (ValueError, IndexError, binascii.Error) as e:
                     logger.debug("Failed to decode multimodal image url: %s", e)
 
-    return ("\n".join(text_parts) if text_parts else None), (images or None)
+    if not parts:
+        return None, None
+
+    return _flatten_parts(parts)
 
 
 @register_genai_provider(GenAIProviderEnum.ollama)
@@ -196,58 +221,46 @@ class OllamaClient(GenAIClient):
         images: list[bytes],
         response_format: dict | None = None,
         enable_thinking: bool = False,
+        image_captions: list[str] | None = None,
     ) -> str | None:
-        """Submit a request to Ollama"""
+        """Submit a request to Ollama through the chat API, the same path the
+        tool-calling chat uses, with image placeholders keeping any captions
+        next to their frames."""
         if self.provider is None:
             logger.warning(
                 "Ollama provider has not been initialized, a description will not be generated. Check your Ollama configuration."
             )
             return None
+
+        content, message_images = _flatten_parts(
+            interleave_images(prompt, images, image_captions)
+        )
+        message: dict[str, Any] = {"role": "user", "content": content}
+
+        if message_images:
+            message["images"] = message_images
+
+        request_params = self._build_request_params(
+            [message], None, None, enable_thinking=enable_thinking
+        )
+
+        if response_format and response_format.get("type") == "json_schema":
+            schema = response_format.get("json_schema", {}).get("schema")
+            if schema:
+                request_params["format"] = self._clean_schema_for_ollama(schema)
+
+        logger.debug(
+            "Ollama chat request: model=%s, prompt_len=%s, image_count=%s, "
+            "has_format=%s, think=%s",
+            self.genai_config.model,
+            len(prompt),
+            len(images),
+            "format" in request_params,
+            request_params.get("think"),
+        )
+
         try:
-            ollama_options = {
-                **self.provider_options,
-                **self.genai_config.runtime_options,
-            }
-            if response_format and response_format.get("type") == "json_schema":
-                schema = response_format.get("json_schema", {}).get("schema")
-                if schema:
-                    ollama_options["format"] = self._clean_schema_for_ollama(schema)
-            if self.supports_toggleable_thinking:
-                ollama_options["think"] = enable_thinking
-            logger.debug(
-                "Ollama generate request: model=%s, prompt_len=%s, image_count=%s, "
-                "has_format=%s, options=%s",
-                self.genai_config.model,
-                len(prompt),
-                len(images) if images else 0,
-                "format" in ollama_options,
-                {k: v for k, v in ollama_options.items() if k != "format"},
-            )
-            result = self.provider.generate(
-                self.genai_config.model,
-                prompt,
-                images=images if images else None,
-                **ollama_options,
-            )
-            logger.debug(
-                "Ollama generate response: done=%s, done_reason=%s, eval_count=%s, "
-                "prompt_eval_count=%s, response_len=%s",
-                result.get("done"),
-                result.get("done_reason"),
-                result.get("eval_count"),
-                result.get("prompt_eval_count"),
-                len(result.get("response", "") or ""),
-            )
-            response_text = str(result["response"]).strip()
-            if not response_text:
-                logger.warning(
-                    "Ollama returned a blank response for model %s (done_reason=%s, "
-                    "eval_count=%s). Check model output, ensure thinking is disabled.",
-                    self.genai_config.model,
-                    result.get("done_reason"),
-                    result.get("eval_count"),
-                )
-            return response_text
+            response = self.provider.chat(**request_params)
         except (
             TimeoutException,
             ResponseError,
@@ -257,6 +270,27 @@ class OllamaClient(GenAIClient):
             logger.warning("Ollama returned an error: %s", str(e))
             return None
 
+        logger.debug(
+            "Ollama chat response: done=%s, done_reason=%s, eval_count=%s, "
+            "prompt_eval_count=%s",
+            response.get("done"),
+            response.get("done_reason"),
+            response.get("eval_count"),
+            response.get("prompt_eval_count"),
+        )
+        response_text = self._message_from_response(response)["content"] or ""
+
+        if not response_text:
+            logger.warning(
+                "Ollama returned a blank response for model %s (done_reason=%s, "
+                "eval_count=%s). Check model output, ensure thinking is disabled.",
+                self.genai_config.model,
+                response.get("done_reason"),
+                response.get("eval_count"),
+            )
+
+        return response_text
+
     def list_models(self) -> list[str]:
         """Return available model names from the Ollama server."""
         client = self.provider
@@ -306,6 +340,8 @@ class OllamaClient(GenAIClient):
             }
             if images:
                 msg_dict["images"] = images
+            elif msg.get("images"):
+                msg_dict["images"] = msg["images"]
             if msg.get("tool_call_id"):
                 msg_dict["tool_call_id"] = msg["tool_call_id"]
             if msg.get("name"):

frigate/genai/plugins/openai.py

@@ -11,9 +11,16 @@ from openai import OpenAI
 
 from frigate.config import GenAIProviderEnum
 from frigate.genai import GenAIClient, register_genai_provider
+from frigate.genai.utils import interleave_images
 
 logger = logging.getLogger(__name__)
 
+# gpt-transcribe replaced the singular `language` field with a `languages` array
+# and rejects a request that sends both. Older transcription models
+# (gpt-4o-transcribe, gpt-4o-mini-transcribe, whisper-1) still take the singular
+# form. https://developers.openai.com/api/docs/guides/speech-to-text
+_LANGUAGES_ARRAY_MODEL_PREFIX = "gpt-transcribe"
+
 
 def _stats_from_openai_usage(usage: Any) -> dict[str, Any] | None:
     """Build a stats dict from an OpenAI-compatible usage object."""
@@ -63,21 +70,21 @@ class OpenAIClient(GenAIClient):
         images: list[bytes],
         response_format: dict | None = None,
         enable_thinking: bool = False,
+        image_captions: list[str] | None = None,
     ) -> str | None:
         """Submit a request to OpenAI."""
-        encoded_images = [base64.b64encode(image).decode("utf-8") for image in images]
-        messages_content: list[dict] = [
-            {
-                "type": "text",
-                "text": prompt,
-            }
-        ]
-        for image in encoded_images:
+        messages_content: list[dict] = []
+        for part in interleave_images(prompt, images, image_captions):
+            if isinstance(part, str):
+                messages_content.append({"type": "text", "text": part})
+                continue
+
+            encoded = base64.b64encode(part).decode("utf-8")
             messages_content.append(
                 {
                     "type": "image_url",
                     "image_url": {
-                        "url": f"data:image/jpeg;base64,{image}",
+                        "url": f"data:image/jpeg;base64,{encoded}",
                         "detail": "low",
                     },
                 }
@@ -133,6 +140,51 @@ class OpenAIClient(GenAIClient):
             logger.warning("OpenAI returned an error: %s", str(e))
             return None
 
+    @property
+    def supports_transcription(self) -> bool:
+        """OpenAI exposes /v1/audio/transcriptions for its speech models."""
+        return True
+
+    def transcribe(
+        self,
+        audio: bytes,
+        language: str | None = None,
+        mime_type: str = "audio/wav",
+    ) -> str | None:
+        """Transcribe audio via the OpenAI audio transcriptions endpoint."""
+        try:
+            # runtime_options are chat-completion parameters; the transcriptions
+            # endpoint rejects unknown fields, so they are deliberately not splatted
+            # in here the way _send() does.
+            request_params: dict[str, Any] = {
+                "model": self.genai_config.model,
+                "file": ("audio.wav", audio, mime_type),
+                "response_format": "text",
+                "timeout": self.timeout,
+            }
+
+            if language:
+                if (
+                    self.genai_config.model.strip()
+                    .lower()
+                    .startswith(_LANGUAGES_ARRAY_MODEL_PREFIX)
+                ):
+                    # not a typed parameter on the SDK method, so it has to ride
+                    # along in extra_body
+                    request_params["extra_body"] = {"languages": [language]}
+                else:
+                    request_params["language"] = language
+
+            result = self.provider.audio.transcriptions.create(**request_params)
+        except (TimeoutException, Exception) as e:
+            logger.warning("OpenAI returned an error: %s", str(e))
+            return None
+
+        # response_format="text" yields a bare string, but some compatible
+        # servers still return the object form
+        text = result if isinstance(result, str) else getattr(result, "text", None)
+        return text.strip() if text else None
+
     def list_models(self) -> list[str]:
         """Return available model IDs from the OpenAI-compatible API."""
         try:

frigate/genai/prompts.py

@@ -59,6 +59,13 @@ def get_review_field_guidelines(response_style: str = "default") -> dict[str, st
     }
 
 
+# Explains the per-frame labels and tracker notes used by the annotated frame
+# mode. Neither the notes nor this guidance say whether repeated detections are
+# the same subject, since the tracking data cannot tell.
+FRAME_ANNOTATION_GUIDANCE = """- Each image below is immediately preceded by a text label giving its frame number and how many seconds into the sequence it was captured. Use these labels to track the order of events and the time between them.
+- Some images below are preceded by notes from the camera's object tracker recording what changed at that point: an object being first detected, starting to move, reversing direction, stopping, or no longer being detected. These notes come from tracking data rather than from the images, and they are reliable. Use them to establish how many distinct activities occur and in what order, and describe every one of them."""
+
+
 def build_review_description_prompt(
     review_data: dict[str, Any],
     thumbnails: list[bytes],
@@ -66,8 +73,13 @@ def build_review_description_prompt(
     preferred_language: str | None,
     activity_context_prompt: str,
     response_style: str = "default",
+    frame_captions: list[str] | None = None,
 ) -> str:
-    """Build the prompt for review activity description generation."""
+    """Build the prompt for review activity description generation.
+
+    When `frame_captions` is set, each caption is sent directly before its
+    image, so the prompt explains that layout.
+    """
 
     def get_concern_prompt() -> str:
         if concerns:
@@ -93,6 +105,7 @@ def build_review_description_prompt(
             return "\n- (No objects detected)"
 
     fields = get_review_field_guidelines(response_style)
+    frame_guidance = f"\n{FRAME_ANNOTATION_GUIDANCE}" if frame_captions else ""
 
     return f"""
 Your task is to analyze a sequence of images taken in chronological order from a security camera.
@@ -130,7 +143,7 @@ Respond with a JSON object matching the provided schema. Field-specific guidance
 ## Sequence Details
 
 - Camera: {review_data["camera"]}
-- Total frames: {len(thumbnails)} (Frame 1 = earliest, Frame {len(thumbnails)} = latest)
+- Total frames: {len(thumbnails)} (Frame 1 = earliest, Frame {len(thumbnails)} = latest){frame_guidance}
 - Activity started at {review_data["start"]} and lasted {review_data["duration"]} seconds
 - Zones involved: {", ".join(review_data["zones"]) if review_data["zones"] else "None"}
 

frigate/genai/utils.py

@@ -7,6 +7,25 @@ from typing import Any
 logger = logging.getLogger(__name__)
 
 
+def interleave_images(
+    prompt: str, images: list[bytes], captions: list[str] | None = None
+) -> list[str | bytes]:
+    """The prompt, then each image preceded by its caption when one is given.
+
+    Providers map the text and image parts onto their own request format, so
+    every provider sends the same order.
+    """
+    parts: list[str | bytes] = [prompt]
+
+    for index, image in enumerate(images):
+        if captions and index < len(captions):
+            parts.append(captions[index])
+
+        parts.append(image)
+
+    return parts
+
+
 def parse_tool_calls_from_message(
     message: dict[str, Any],
 ) -> list[dict[str, Any]] | None:

frigate/log.py

@@ -144,6 +144,14 @@ class LogPipe(threading.Thread):
         self.pipeReader.close()
 
     def dump(self) -> None:
+        if not self.deque:
+            return
+
+        self.logger.log(
+            self.level,
+            "The following ffmpeg logs include the last 100 lines prior to exit.",
+        )
+
         while len(self.deque) > 0:
             self.logger.log(self.level, self.deque.popleft())
 

frigate/motion/improved_motion.py

@@ -188,8 +188,12 @@ class ImprovedMotionDetector(MotionDetector):
             self.config.skip_motion_threshold is not None
             and pct_motion > self.config.skip_motion_threshold
         ):
-            # force a recalibration so we transition to the new background
+            # recalibrate so we transition to the new background. the frame
+            # still has to be blended in here, otherwise the background stays
+            # frozen and every subsequent frame skips as well
             self.calibrating = True
+            cv2.accumulateWeighted(resized_frame, self.avg_frame, 0.2)
+            self.motion_frame_count = 0
             return []
 
         # once the motion is less than 5% and the number of contours is < 4, assume its calibrated

frigate/ptz/onvif.py

@@ -1120,6 +1120,18 @@ class OnvifController:
                     f"Camera {camera_name} is still in ONVIF 'MOVING' status."
                 )
 
+    async def _shutdown(self) -> None:
+        """Close the camera sessions and cancel the tasks running on the loop."""
+        for cam_name in list(self.cams):
+            await self._close_camera(cam_name)
+
+        tasks = [t for t in asyncio.all_tasks() if t is not asyncio.current_task()]
+
+        for task in tasks:
+            task.cancel()
+
+        await asyncio.gather(*tasks, return_exceptions=True)
+
     def close(self) -> None:
         """Gracefully shut down the ONVIF controller."""
         if not hasattr(self, "loop") or self.loop.is_closed():
@@ -1127,6 +1139,16 @@ class OnvifController:
             return
 
         logger.info("Exiting ONVIF controller...")
+
+        # anything left open here is garbage collected during interpreter
+        # shutdown, where its warnings can no longer be logged cleanly
+        try:
+            asyncio.run_coroutine_threadsafe(self._shutdown(), self.loop).result(
+                timeout=5
+            )
+        except TimeoutError:
+            logger.debug("Timed out closing ONVIF sessions")
+
         self.config_subscriber.stop()
 
         def stop_and_cleanup():

frigate/test/test_audio_spawn.py

@@ -0,0 +1,90 @@
+"""Tests for audio maintainer spawning on runtime camera adds."""
+
+import threading
+import unittest
+from unittest.mock import MagicMock, patch
+
+from frigate.config import FrigateConfig
+from frigate.events.audio import AudioProcessor
+
+
+def _build_config() -> FrigateConfig:
+    return FrigateConfig(
+        **{
+            "mqtt": {"host": "mqtt"},
+            "cameras": {
+                "front_door": {
+                    "ffmpeg": {
+                        "inputs": [
+                            {
+                                "path": "rtsp://10.0.0.1:554/video",
+                                "roles": ["detect", "audio"],
+                            }
+                        ]
+                    },
+                    "audio": {"enabled": True},
+                }
+            },
+        }
+    )
+
+
+class TestAudioSpawnIfNeeded(unittest.TestCase):
+    def _processor(self, config: FrigateConfig, camera_metrics: dict) -> AudioProcessor:
+        processor = AudioProcessor.__new__(AudioProcessor)
+        processor.config = config
+        processor.camera_metrics = camera_metrics
+        processor.embeddings_metrics = MagicMock()
+        processor.audio_threads = {}
+        processor.transcription_model_runner = None
+        processor.genai_manager = None
+        processor.stop_event = threading.Event()
+        processor.logger = MagicMock()
+        return processor
+
+    def test_skips_camera_whose_metrics_have_not_been_created_yet(self):
+        """The camera maintainer creates metrics on its own poll of the add
+        update, so the audio processor can see the camera first."""
+        config = _build_config()
+        processor = self._processor(config, {})
+
+        with patch("frigate.events.audio.AudioEventMaintainer") as maintainer:
+            processor.spawn_if_needed(config.cameras["front_door"])
+
+        maintainer.assert_not_called()
+        assert processor.audio_threads == {}
+
+    def test_spawns_once_metrics_exist_and_passes_the_metrics_object(self):
+        config = _build_config()
+        metrics = object()
+        processor = self._processor(config, {"front_door": metrics})
+
+        with patch("frigate.events.audio.AudioEventMaintainer") as maintainer:
+            processor.spawn_if_needed(config.cameras["front_door"])
+
+        assert maintainer.call_args.args[2] is metrics
+        assert "front_door" in processor.audio_threads
+        maintainer.return_value.start.assert_called_once()
+
+    def test_does_not_respawn_for_a_camera_already_running(self):
+        config = _build_config()
+        processor = self._processor(config, {"front_door": object()})
+        processor.audio_threads["front_door"] = MagicMock()
+
+        with patch("frigate.events.audio.AudioEventMaintainer") as maintainer:
+            processor.spawn_if_needed(config.cameras["front_door"])
+
+        maintainer.assert_not_called()
+
+    def test_skips_camera_whose_ffmpeg_update_has_not_arrived_yet(self):
+        """The add update carries the camera before its ffmpeg inputs are
+        applied, so the audio role can be briefly missing."""
+        config = _build_config()
+        camera = config.cameras["front_door"]
+        camera.ffmpeg.inputs[0].roles = ["detect"]
+        processor = self._processor(config, {"front_door": object()})
+
+        with patch("frigate.events.audio.AudioEventMaintainer") as maintainer:
+            processor.spawn_if_needed(camera)
+
+        maintainer.assert_not_called()

frigate/test/test_audio_transcription_genai.py

@@ -0,0 +1,286 @@
+"""Config validation and the historical path for the audio_transcription GenAI backend."""
+
+import unittest
+from copy import deepcopy
+from unittest.mock import MagicMock, patch
+
+from pydantic import ValidationError
+
+from frigate.config import FrigateConfig
+from frigate.config.camera.genai import GenAIConfig, GenAIRoleEnum
+from frigate.config.classification import AudioTranscriptionModelEnum
+from frigate.const import UPDATE_EVENT_DESCRIPTION
+from frigate.data_processing.post.audio_transcription import (
+    AudioTranscriptionPostProcessor,
+)
+from frigate.data_processing.types import PostProcessDataEnum
+
+
+class TestAudioTranscriptionGenAIConfig(unittest.TestCase):
+    def setUp(self):
+        self.base = {
+            "mqtt": {"host": "mqtt"},
+            "cameras": {
+                "back": {
+                    "ffmpeg": {
+                        "inputs": [
+                            {
+                                "path": "rtsp://10.0.0.1:554/video",
+                                "roles": ["detect", "audio"],
+                            }
+                        ]
+                    },
+                    "detect": {"height": 1080, "width": 1920, "fps": 5},
+                    "audio": {"enabled": True},
+                }
+            },
+        }
+
+    def _config(self, **overrides) -> dict:
+        config = deepcopy(self.base)
+        config.update(deepcopy(overrides))
+        return config
+
+    def _provider(self, roles: list[str]) -> dict:
+        return {
+            "whisper_cloud": {
+                "provider": "openai",
+                "model": "gpt-4o-transcribe",
+                "api_key": "k",
+                "roles": roles,
+            }
+        }
+
+    def test_default_model_is_whisper_enum(self):
+        config = FrigateConfig(**self._config())
+        self.assertEqual(
+            config.audio_transcription.model, AudioTranscriptionModelEnum.whisper
+        )
+
+    def test_whisper_string_coerces_to_enum(self):
+        config = FrigateConfig(
+            **self._config(audio_transcription={"enabled": True, "model": "whisper"})
+        )
+        self.assertIsInstance(
+            config.audio_transcription.model, AudioTranscriptionModelEnum
+        )
+
+    def test_provider_name_stays_a_string(self):
+        config = FrigateConfig(
+            **self._config(
+                genai=self._provider(["transcribe"]),
+                audio_transcription={"enabled": True, "model": "whisper_cloud"},
+            )
+        )
+        self.assertNotIsInstance(
+            config.audio_transcription.model, AudioTranscriptionModelEnum
+        )
+        self.assertEqual(config.audio_transcription.model, "whisper_cloud")
+
+    def test_unspecified_model_falls_back_to_whisper(self):
+        """An empty value must not read as a GenAI provider that resolves to no client."""
+        for value in (None, "", "   "):
+            with self.subTest(repr(value)):
+                config = FrigateConfig(
+                    **self._config(
+                        audio_transcription={"enabled": True, "model": value}
+                    )
+                )
+                self.assertIs(
+                    config.audio_transcription.model,
+                    AudioTranscriptionModelEnum.whisper,
+                )
+
+    def test_missing_genai_key_raises(self):
+        with self.assertRaises(ValidationError) as ctx:
+            FrigateConfig(
+                **self._config(
+                    audio_transcription={"enabled": True, "model": "nope"},
+                )
+            )
+        self.assertIn("is not a valid GenAI config key", str(ctx.exception))
+
+    def test_provider_without_role_raises(self):
+        with self.assertRaises(ValidationError) as ctx:
+            FrigateConfig(
+                **self._config(
+                    genai=self._provider(["descriptions"]),
+                    audio_transcription={"enabled": True, "model": "whisper_cloud"},
+                )
+            )
+        self.assertIn("must have 'transcribe' in its roles", str(ctx.exception))
+
+    def test_global_off_camera_on_still_validates(self):
+        """Global-off/camera-on is a supported deployment and must not skip the check."""
+        config = self._config(
+            genai=self._provider(["descriptions"]),
+            audio_transcription={"enabled": False, "model": "whisper_cloud"},
+        )
+        config["cameras"]["back"]["audio_transcription"] = {"enabled": True}
+
+        with self.assertRaises(ValidationError) as ctx:
+            FrigateConfig(**config)
+
+        self.assertIn("must have 'transcribe' in its roles", str(ctx.exception))
+
+    def test_disabled_transcription_skips_validation(self):
+        config = FrigateConfig(
+            **self._config(audio_transcription={"enabled": False, "model": "nope"})
+        )
+        self.assertEqual(config.audio_transcription.model, "nope")
+
+    def test_camera_level_model_is_rejected(self):
+        config = self._config()
+        config["cameras"]["back"]["audio_transcription"] = {
+            "enabled": True,
+            "model": "whisper",
+        }
+
+        with self.assertRaises(ValidationError):
+            FrigateConfig(**config)
+
+    def test_default_roles_do_not_include_transcribe(self):
+        """Backward compatibility: existing providers must not silently claim it."""
+        genai = GenAIConfig(provider="openai", model="gpt-4o")
+        self.assertNotIn(GenAIRoleEnum.transcribe, genai.roles)
+
+    def test_two_providers_claiming_transcribe_raises(self):
+        genai = self._provider(["transcribe"])
+        genai["other"] = {
+            "provider": "gemini",
+            "model": "gemini-2.0-flash",
+            "api_key": "k",
+            "roles": ["transcribe"],
+        }
+
+        with self.assertRaises(ValidationError) as ctx:
+            FrigateConfig(
+                **self._config(
+                    genai=genai,
+                    audio_transcription={"enabled": True, "model": "whisper_cloud"},
+                )
+            )
+
+        self.assertIn("each role must have", str(ctx.exception))
+
+    def test_transcribe_rejected_on_provider_without_audio_input(self):
+        with self.assertRaises(ValidationError) as ctx:
+            GenAIConfig(provider="ollama", model="llava", roles=["transcribe"])
+
+        self.assertIn("does not support audio input", str(ctx.exception))
+
+
+class TestAudioTranscriptionPostProcessorGenAI(unittest.TestCase):
+    """The recorded-speech path must reach the provider and skip the local model."""
+
+    def setUp(self):
+        self.config = FrigateConfig(
+            **{
+                "mqtt": {"host": "mqtt"},
+                "genai": {
+                    "whisper_cloud": {
+                        "provider": "openai",
+                        "model": "gpt-4o-transcribe",
+                        "api_key": "k",
+                        "roles": ["transcribe"],
+                    }
+                },
+                "audio_transcription": {
+                    "enabled": True,
+                    "model": "whisper_cloud",
+                    "language": "en",
+                },
+                "cameras": {
+                    "back": {
+                        "ffmpeg": {
+                            "inputs": [
+                                {
+                                    "path": "rtsp://10.0.0.1:554/video",
+                                    "roles": ["detect", "audio"],
+                                }
+                            ]
+                        },
+                        "detect": {"height": 1080, "width": 1920, "fps": 5},
+                        "audio": {"enabled": True},
+                    }
+                },
+            }
+        )
+
+        self.client = MagicMock()
+        self.client.transcribe.return_value = "recorded speech"
+        self.manager = MagicMock()
+        self.manager.transcribe_client = self.client
+        self.requestor = MagicMock()
+
+        self.processor = AudioTranscriptionPostProcessor(
+            self.config,
+            self.requestor,
+            MagicMock(),
+            MagicMock(),
+            self.manager,
+        )
+
+    def _process(self):
+        self.processor.process_data(
+            {
+                "event_id": "1234.5-abc",
+                "camera": "back",
+                "event": {
+                    "id": "1234.5-abc",
+                    "camera": "back",
+                    "start_time": 100.0,
+                    "end_time": 110.0,
+                    "data": {},
+                },
+            },
+            PostProcessDataEnum.tracked_object,
+        )
+
+    def test_local_recognizer_is_never_built(self):
+        self.assertTrue(self.processor._use_genai)
+        self.assertIsNone(self.processor.recognizer)
+
+    def test_audio_bytes_and_language_reach_the_client(self):
+        with patch(
+            "frigate.data_processing.post.audio_transcription.get_audio_from_recording",
+            return_value=b"RIFF....WAVE",
+        ):
+            self._process()
+
+        self.client.transcribe.assert_called_once()
+        self.assertEqual(self.client.transcribe.call_args.args[0], b"RIFF....WAVE")
+        self.assertEqual(self.client.transcribe.call_args.kwargs["language"], "en")
+
+    def test_transcript_is_published_as_the_description(self):
+        with patch(
+            "frigate.data_processing.post.audio_transcription.get_audio_from_recording",
+            return_value=b"RIFF....WAVE",
+        ):
+            self._process()
+
+        topics = [call.args[0] for call in self.requestor.send_data.call_args_list]
+        self.assertIn(UPDATE_EVENT_DESCRIPTION, topics)
+
+        payload = next(
+            call.args[1]
+            for call in self.requestor.send_data.call_args_list
+            if call.args[0] == UPDATE_EVENT_DESCRIPTION
+        )
+        self.assertEqual(payload["description"], "recorded speech")
+        self.assertEqual(payload["id"], "1234.5-abc")
+
+    def test_missing_client_publishes_nothing(self):
+        self.manager.transcribe_client = None
+
+        with patch(
+            "frigate.data_processing.post.audio_transcription.get_audio_from_recording",
+            return_value=b"RIFF....WAVE",
+        ):
+            self._process()
+
+        self.requestor.send_data.assert_not_called()
+
+
+if __name__ == "__main__":
+    unittest.main()

frigate/test/test_audio_transcription_live_buffer.py

@@ -0,0 +1,288 @@
+"""Live GenAI transcription: sliding overlapped windows and their lifecycle."""
+
+import io
+import threading
+import unittest
+import wave
+from unittest.mock import MagicMock, patch
+
+import numpy as np
+
+from frigate.config import FrigateConfig
+from frigate.const import AUDIO_DURATION, AUDIO_SAMPLE_RATE
+from frigate.data_processing.real_time.audio_transcription import (
+    GENAI_WINDOW_CHUNKS,
+    AudioTranscriptionRealTimeProcessor,
+)
+
+CHUNK_SAMPLES = int(round(AUDIO_DURATION * AUDIO_SAMPLE_RATE))
+
+
+def _chunk(amplitude: int) -> np.ndarray:
+    """One audio-detector-sized chunk of int16 samples at a constant amplitude."""
+    return np.full(CHUNK_SAMPLES, amplitude, dtype=np.int16)
+
+
+class TestLiveGenAITranscription(unittest.TestCase):
+    def setUp(self):
+        self.config = FrigateConfig(
+            **{
+                "mqtt": {"host": "mqtt"},
+                "genai": {
+                    "whisper_cloud": {
+                        "provider": "openai",
+                        "model": "gpt-4o-transcribe",
+                        "api_key": "k",
+                        "roles": ["transcribe"],
+                    }
+                },
+                "audio_transcription": {
+                    "enabled": True,
+                    "model": "whisper_cloud",
+                    "language": "en",
+                },
+                "cameras": {
+                    "back": {
+                        "ffmpeg": {
+                            "inputs": [
+                                {
+                                    "path": "rtsp://10.0.0.1:554/video",
+                                    "roles": ["detect", "audio"],
+                                }
+                            ]
+                        },
+                        "detect": {"height": 1080, "width": 1920, "fps": 5},
+                        "audio": {"enabled": True},
+                    }
+                },
+            }
+        )
+
+        self.client = MagicMock()
+        self.client.transcribe.return_value = "hello"
+        self.manager = MagicMock()
+        self.manager.transcribe_client = self.client
+        self.requestor = MagicMock()
+
+        self.processor = AudioTranscriptionRealTimeProcessor(
+            config=self.config,
+            camera_config=self.config.cameras["back"],
+            requestor=self.requestor,
+            model_runner=None,
+            metrics=MagicMock(),
+            stop_event=threading.Event(),
+            genai_manager=self.manager,
+        )
+
+    def _feed(self, chunk: np.ndarray):
+        return (
+            self.processor._AudioTranscriptionRealTimeProcessor__process_audio_stream(
+                chunk
+            )
+        )
+
+    def _sent_wav(self, call_index: int) -> wave.Wave_read:
+        payload = self.client.transcribe.call_args_list[call_index].args[0]
+        return wave.open(io.BytesIO(payload), "rb")
+
+    def test_uses_genai_path(self):
+        self.assertTrue(self.processor._use_genai)
+
+    def test_first_chunk_does_not_transcribe(self):
+        self.assertIsNone(self._feed(_chunk(4000)))
+        self.client.transcribe.assert_not_called()
+
+    def test_full_window_transcribes_two_chunks(self):
+        self._feed(_chunk(4000))
+        result = self._feed(_chunk(4000))
+
+        self.assertEqual(result, ("hello", False))
+        self.client.transcribe.assert_called_once()
+        self.assertEqual(
+            self.client.transcribe.call_args.kwargs["language"],
+            "en",
+        )
+
+        with self._sent_wav(0) as wav:
+            self.assertEqual(wav.getnchannels(), 1)
+            self.assertEqual(wav.getsampwidth(), 2)
+            self.assertEqual(wav.getframerate(), AUDIO_SAMPLE_RATE)
+            self.assertEqual(wav.getnframes(), CHUNK_SAMPLES * GENAI_WINDOW_CHUNKS)
+
+    def test_window_slides_with_overlap(self):
+        """The third chunk's window is chunks 2+3, not 3 alone and not 1+2+3."""
+        self.client.transcribe.side_effect = ["one two", "two three"]
+
+        self._feed(_chunk(1000))
+        self._feed(_chunk(2000))
+        result = self._feed(_chunk(3000))
+
+        self.assertEqual(self.client.transcribe.call_count, 2)
+
+        with self._sent_wav(1) as wav:
+            self.assertEqual(wav.getnframes(), CHUNK_SAMPLES * GENAI_WINDOW_CHUNKS)
+            samples = np.frombuffer(wav.readframes(wav.getnframes()), dtype=np.int16)
+
+        self.assertEqual(samples[0], 2000)
+        self.assertEqual(samples[-1], 3000)
+
+        # the shared "two" appears once
+        self.assertEqual(result, ("one two three", False))
+
+    def test_silent_window_is_not_uploaded(self):
+        self._feed(_chunk(0))
+        self.assertIsNone(self._feed(_chunk(0)))
+        self.client.transcribe.assert_not_called()
+
+    def test_silent_window_ends_a_pending_utterance(self):
+        self._feed(_chunk(4000))
+        self._feed(_chunk(4000))
+
+        # the gate covers the whole window, so it takes GENAI_WINDOW_CHUNKS
+        # silent chunks to push the last speech out of it
+        self._feed(_chunk(0))
+
+        self.assertEqual(self._feed(_chunk(0)), ("hello", True))
+
+    def test_empty_transcript_ends_a_pending_utterance(self):
+        self.client.transcribe.side_effect = ["hello", ""]
+
+        self._feed(_chunk(4000))
+        self._feed(_chunk(4000))
+
+        self.assertEqual(self._feed(_chunk(4000)), ("hello", True))
+
+    def test_empty_transcript_with_nothing_pending_returns_none(self):
+        self.client.transcribe.return_value = ""
+
+        self._feed(_chunk(4000))
+        self.assertIsNone(self._feed(_chunk(4000)))
+
+    def test_reset_clears_committed_text_and_window(self):
+        self._feed(_chunk(4000))
+        self._feed(_chunk(4000))
+
+        self.processor.reset()
+
+        self.assertEqual(self.processor._genai_committed, "")
+        self.assertEqual(len(self.processor._genai_window), 0)
+
+        # a fresh window is required again before the next request
+        self.client.transcribe.reset_mock()
+        self._feed(_chunk(4000))
+        self.client.transcribe.assert_not_called()
+
+    def test_check_unload_model_clears_once_then_is_idempotent(self):
+        self._feed(_chunk(4000))
+        self._feed(_chunk(4000))
+
+        self.processor.check_unload_model()
+
+        self.requestor.send_data.assert_called_once_with("back/audio/transcription", "")
+        self.assertEqual(self.processor._genai_committed, "")
+        self.assertEqual(len(self.processor._genai_window), 0)
+
+        self.processor.check_unload_model()
+        self.requestor.send_data.assert_called_once()
+
+    def test_build_recognizer_never_loads_a_local_model(self):
+        with patch(
+            "frigate.data_processing.real_time.audio_transcription.FasterWhisperASR"
+        ) as whisper:
+            self.processor._AudioTranscriptionRealTimeProcessor__build_recognizer()
+
+        whisper.assert_not_called()
+        self.assertIsNone(self.processor.stream)
+
+    def test_clear_audio_recognizer_request_only_resets(self):
+        self._feed(_chunk(4000))
+        self._feed(_chunk(4000))
+
+        with patch.object(
+            self.processor,
+            "_AudioTranscriptionRealTimeProcessor__build_recognizer",
+        ) as build:
+            result = self.processor.handle_request("clear_audio_recognizer", {})
+
+        build.assert_not_called()
+        self.assertTrue(result["success"])
+        self.assertEqual(self.processor._genai_committed, "")
+
+    def test_missing_client_logs_and_returns_none(self):
+        self.manager.transcribe_client = None
+
+        self.assertIsNone(self._feed(_chunk(4000)))
+        self.assertIsNone(self._feed(_chunk(4000)))
+
+    def test_dropped_audio_discards_the_buffered_window(self):
+        """A gap in the stream must not be spliced into a single window.
+
+        Dropping a queued chunk leaves the next one non-adjacent to what is
+        buffered, so concatenating them would hand the provider audio with a
+        hole in it and break the 50% overlap the stitcher relies on.
+        """
+        self._feed(_chunk(4000))
+        self.assertEqual(len(self.processor._genai_window), 1)
+
+        # the producer discards a chunk while the consumer is blocked
+        self.processor._audio_dropped.set()
+
+        self._feed(_chunk(5000))
+
+        # the buffered chunk was discarded, so this one starts a fresh window
+        self.assertEqual(len(self.processor._genai_window), 1)
+        self.client.transcribe.assert_not_called()
+
+        # and the window that does go out holds only contiguous audio
+        self._feed(_chunk(5000))
+        self.client.transcribe.assert_called_once()
+
+        with self._sent_wav(0) as wav:
+            samples = np.frombuffer(wav.readframes(wav.getnframes()), dtype=np.int16)
+
+        self.assertEqual(wav.getnframes(), CHUNK_SAMPLES * GENAI_WINDOW_CHUNKS)
+        self.assertTrue((samples == 5000).all(), "window spliced across the gap")
+
+    def test_dropped_audio_ends_a_pending_utterance(self):
+        """Committed text cannot be stitched across missing speech."""
+        self._feed(_chunk(4000))
+        self._feed(_chunk(4000))
+        self.assertEqual(self.processor._genai_committed, "hello")
+
+        self.processor._audio_dropped.set()
+
+        self.assertEqual(self._feed(_chunk(4000)), ("hello", True))
+        self.assertEqual(len(self.processor._genai_window), 0)
+
+    def test_drop_flag_is_consumed_once(self):
+        self.processor._audio_dropped.set()
+        self._feed(_chunk(4000))
+
+        self.assertFalse(self.processor._audio_dropped.is_set())
+
+    def test_full_queue_flags_a_drop(self):
+        for i in range(self.processor.audio_queue.maxsize + 1):
+            self.processor.process_audio({"id": "back_audio"}, _chunk(i + 1))
+
+        self.assertTrue(self.processor._audio_dropped.is_set())
+
+    def test_queue_is_bounded_and_drops_oldest(self):
+        maxsize = self.processor.audio_queue.maxsize
+        self.assertGreater(maxsize, 0)
+
+        for i in range(maxsize + 5):
+            self.processor.process_audio({"id": "back_audio"}, _chunk(i + 1))
+
+        self.assertEqual(self.processor.audio_queue.qsize(), maxsize)
+
+        # the newest chunk survived, the oldest did not
+        remaining = []
+        while not self.processor.audio_queue.empty():
+            remaining.append(self.processor.audio_queue.get_nowait()[1][0])
+
+        self.assertEqual(remaining[-1], maxsize + 5)
+        self.assertNotIn(1, remaining)
+
+
+if __name__ == "__main__":
+    unittest.main()

frigate/test/test_audio_util.py

@@ -0,0 +1,176 @@
+"""Tests for the WAV helpers and transcript stitcher in frigate.util.audio."""
+
+import io
+import struct
+import unittest
+import wave
+
+import numpy as np
+
+from frigate.const import AUDIO_SAMPLE_RATE
+from frigate.util.audio import fix_wav_header, pcm16_to_wav, stitch_transcripts
+
+
+def _wav(samples: np.ndarray, sample_rate: int = AUDIO_SAMPLE_RATE) -> bytes:
+    buffer = io.BytesIO()
+
+    with wave.open(buffer, "wb") as out:
+        out.setnchannels(1)
+        out.setsampwidth(2)
+        out.setframerate(sample_rate)
+        out.writeframes(samples.tobytes())
+
+    return buffer.getvalue()
+
+
+class TestPcm16ToWav(unittest.TestCase):
+    def test_round_trips_through_wave(self):
+        samples = np.arange(-1000, 1000, dtype=np.int16)
+
+        with wave.open(io.BytesIO(pcm16_to_wav(samples)), "rb") as wav:
+            self.assertEqual(wav.getnchannels(), 1)
+            self.assertEqual(wav.getsampwidth(), 2)
+            self.assertEqual(wav.getframerate(), AUDIO_SAMPLE_RATE)
+            self.assertEqual(wav.getnframes(), samples.size)
+            decoded = np.frombuffer(wav.readframes(wav.getnframes()), dtype=np.int16)
+
+        np.testing.assert_array_equal(decoded, samples)
+
+    def test_casts_non_int16_input(self):
+        samples = np.array([0.0, 100.0, -100.0], dtype=np.float32)
+
+        with wave.open(io.BytesIO(pcm16_to_wav(samples)), "rb") as wav:
+            decoded = np.frombuffer(wav.readframes(wav.getnframes()), dtype=np.int16)
+
+        np.testing.assert_array_equal(decoded, np.array([0, 100, -100], np.int16))
+
+    def test_honors_sample_rate(self):
+        with wave.open(
+            io.BytesIO(pcm16_to_wav(np.zeros(4, np.int16), 8000)), "rb"
+        ) as w:
+            self.assertEqual(w.getframerate(), 8000)
+
+
+class TestFixWavHeader(unittest.TestCase):
+    def test_rewrites_placeholder_sizes(self):
+        samples = np.arange(64, dtype=np.int16)
+        data = bytearray(_wav(samples))
+
+        # ffmpeg piping to non-seekable stdout leaves both sizes unpatched
+        struct.pack_into("<I", data, 4, 0xFFFFFFFF)
+        data_offset = data.index(b"data")
+        struct.pack_into("<I", data, data_offset + 4, 0xFFFFFFFF)
+
+        fixed = fix_wav_header(bytes(data))
+
+        self.assertEqual(struct.unpack_from("<I", fixed, 4)[0], len(fixed) - 8)
+        self.assertEqual(
+            struct.unpack_from("<I", fixed, data_offset + 4)[0],
+            len(fixed) - (data_offset + 8),
+        )
+
+        with wave.open(io.BytesIO(fixed), "rb") as wav:
+            self.assertEqual(wav.getnframes(), samples.size)
+
+    def test_leaves_a_well_formed_header_alone(self):
+        data = _wav(np.arange(32, dtype=np.int16))
+        self.assertEqual(fix_wav_header(data), data)
+
+    def test_non_riff_payload_passes_through(self):
+        self.assertEqual(fix_wav_header(b"not a wav"), b"not a wav")
+        self.assertEqual(fix_wav_header(b""), b"")
+
+
+class TestStitchTranscripts(unittest.TestCase):
+    def test_table(self):
+        cases = [
+            # (committed, incoming, expected, description)
+            (
+                "the quick brown",
+                "brown fox jumps",
+                "the quick brown fox jumps",
+                "one word",
+            ),
+            (
+                "and then the quick brown",
+                "the quick brown fox",
+                "and then the quick brown fox",
+                "multi word",
+            ),
+            (
+                "hello there",
+                "general kenobi",
+                "hello there general kenobi",
+                "no overlap",
+            ),
+            ("the quick brown fox", "brown fox", "the quick brown fox", "contained"),
+            ("", "first words", "first words", "empty committed"),
+            ("already here", "", "already here", "empty incoming"),
+            (
+                "  spaced   out  ",
+                "out again",
+                "spaced out again",
+                "whitespace normalized",
+            ),
+        ]
+
+        for committed, incoming, expected, description in cases:
+            with self.subTest(description):
+                self.assertEqual(stitch_transcripts(committed, incoming), expected)
+
+    def test_overlap_found_mid_window(self):
+        """The shared run is rarely at the start of the new window.
+
+        The provider re-transcribes the overlapping audio independently and
+        often renders its first word differently, so anchoring the match to the
+        start of the incoming window duplicates the whole phrase.
+        """
+        self.assertEqual(
+            stitch_transcripts(
+                "this is just gonna be a fun time", "It's gonna be a fun time."
+            ),
+            "this is just gonna be a fun time",
+        )
+
+    def test_overlap_longer_than_five_words(self):
+        """The cap is bounded by window duration, not by the old 5-word n-gram."""
+        self.assertEqual(
+            stitch_transcripts(
+                "well anyway one two three four five six",
+                "one two three four five six seven",
+            ),
+            "well anyway one two three four five six seven",
+        )
+
+    def test_repeated_phrase_keeps_its_second_utterance(self):
+        """Preferring the earliest match is what protects a real repeat."""
+        self.assertEqual(
+            stitch_transcripts("a b c fun time", "fun time fun time"),
+            "a b c fun time fun time",
+        )
+
+    def test_window_wholly_repeating_the_tail_is_dropped(self):
+        """The accepted trade-off: an entirely redundant window adds nothing."""
+        self.assertEqual(stitch_transcripts("go go go", "go go go"), "go go go")
+
+    def test_revises_a_mistranscribed_tail(self):
+        """A wrong last word would otherwise block every alignment.
+
+        Those words came from the newest audio, which the next window re-covers,
+        so replacing them is better than duplicating the phrase behind them.
+        """
+        self.assertEqual(
+            stitch_transcripts("Yeah. this is Jessica.", "This is just gonna be fun."),
+            "Yeah. this is just gonna be fun.",
+        )
+
+    def test_revision_needs_more_than_one_shared_word(self):
+        """A revision deletes published text, so it takes real evidence."""
+        self.assertEqual(
+            stitch_transcripts("the cat sat on a mat", "a dog barked"),
+            "the cat sat on a mat a dog barked",
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

frigate/test/test_builtin.py

@@ -36,6 +36,31 @@ class TestEventsPerSecond(unittest.TestCase):
             clock[0] += 100.0
             self.assertEqual(eps.eps(), 0.0)
 
+    def test_burst_after_start_is_not_divided_by_a_tiny_window(self) -> None:
+        eps = EventsPerSecond(last_n_seconds=10)
+        clock = [1000.0]
+        with patch("frigate.util.builtin.time.monotonic", side_effect=lambda: clock[0]):
+            eps.start()
+            # eleven buffered frames arrive within 100 ms of starting
+            for _ in range(11):
+                clock[0] += 0.01
+                eps.update()
+            # 11 events over less than a second is at most 11 per second
+            self.assertLessEqual(eps.eps(), 11.0)
+
+    def test_subsecond_window_keeps_its_rate(self) -> None:
+        eps = EventsPerSecond(last_n_seconds=0.5)
+        clock = [1000.0]
+        with patch("frigate.util.builtin.time.monotonic", side_effect=lambda: clock[0]):
+            eps.start()
+            # twenty events per second for two seconds
+            for _ in range(40):
+                clock[0] += 0.05
+                eps.update()
+            # read between events, so none sits exactly on the window edge
+            clock[0] += 0.01
+            self.assertAlmostEqual(eps.eps(), 20.0)
+
 
 if __name__ == "__main__":
     unittest.main()

frigate/test/test_deepx.py

@@ -0,0 +1,896 @@
+"""Tests for the DEEPX detector."""
+
+import json
+import os
+import struct
+import sys
+import tempfile
+import unittest
+from unittest.mock import MagicMock, patch
+
+import numpy as np
+from pydantic import ValidationError
+
+from frigate.detectors.detector_config import ModelConfig, ModelTypeEnum
+from frigate.detectors.device import (
+    DeviceParseError,
+    build_detector_config,
+    parse_device,
+)
+from frigate.detectors.plugins.deepx import (
+    DEEPX_MANIFEST,
+    DXRT_VERSION,
+    PPU_RECORD_SIZE,
+    DeepxDetector,
+    DeepxDetectorConfig,
+    PpuLayout,
+    YoloLayout,
+    class_count,
+    decode_raw_anchor,
+    decode_raw_nms_in_head,
+    infer_yolo_layout,
+    read_ppu_layout,
+    resolve_device,
+    validate_yolox_outputs,
+)
+
+
+def model_with_type(model_type) -> ModelConfig:
+    return ModelConfig(
+        model_type=model_type,
+        labelmap_path=None,
+        labelmap={79: "toothbrush"},
+        width=640,
+        height=640,
+    )
+
+
+def build_ppu_record(box, score=0.9, label=0, grid=(7, 9, 2, 2)) -> np.ndarray:
+    record = np.zeros(PPU_RECORD_SIZE, dtype=np.uint8)
+    record[0:16] = np.array(box, dtype=np.float32).view(np.uint8)
+    record[16:20] = grid
+    record[20:24] = np.array([score], dtype=np.float32).view(np.uint8)
+    record[24:28] = np.array([label], dtype=np.uint32).view(np.uint8)
+    return record.reshape(1, 1, PPU_RECORD_SIZE)
+
+
+# compile_config.ppu as DX-COM writes it for the two head kinds
+ANCHOR_BASED_PPU = {"type": 0, "num_classes": 80, "activation": "Sigmoid"}
+ANCHOR_FREE_PPU = {"type": 1, "num_classes": 80}
+
+PPU_BBOX_NODE = "/head/Mul_2"
+
+# (grid_w, grid_h, entries) per scale, finest first; the grids give the
+# strides at a 640 input
+THREE_SCALE_ANCHORS = [(80, 80, 3), (40, 40, 3), (20, 20, 3)]
+TWO_SCALE_ANCHORS = [(40, 40, 3), (20, 20, 3)]
+FOUR_SCALE_ANCHORS = [(160, 160, 3), (80, 80, 3), (40, 40, 3), (20, 20, 3)]
+THREE_SCALE_FREE = [(80, 80, 1), (40, 40, 1), (20, 20, 1)]
+ONE_SCALE_FREE = [(100, 84, 1)]
+
+
+def proto_varint(value: int) -> bytes:
+    out = bytearray()
+    while True:
+        byte = value & 0x7F
+        value >>= 7
+        out.append(byte | (0x80 if value else 0))
+        if not value:
+            return bytes(out)
+
+
+def proto_bytes(field: int, payload: bytes) -> bytes:
+    return proto_varint(field << 3 | 2) + proto_varint(len(payload)) + payload
+
+
+def proto_number(field: int, value: int) -> bytes:
+    return proto_varint(field << 3) + proto_varint(value)
+
+
+def onnx_node(op_type: str, name: str, inputs: list, outputs: list) -> bytes:
+    body = b"".join(proto_bytes(1, tensor.encode()) for tensor in inputs)
+    body += b"".join(proto_bytes(2, tensor.encode()) for tensor in outputs)
+    return body + proto_bytes(3, name.encode()) + proto_bytes(4, op_type.encode())
+
+
+def onnx_box_graph(box_format: str) -> bytes:
+    if box_format == "broken":
+        return proto_varint(1 << 3 | 3)
+
+    unnamed = box_format == "unnamed"
+
+    def graph_node(op_type: str, name: str, inputs: list, outputs: list) -> bytes:
+        return onnx_node(op_type, "" if unnamed else name, inputs, outputs)
+
+    nodes = [
+        graph_node("Split", "dfl_split", ["dfl", "sizes"], ["lt", "rb"]),
+        graph_node("Sub", "corner_min", ["anchors", "lt"], ["x1y1"]),
+        graph_node("Add", "corner_max", ["rb", "anchors"], ["x2y2"]),
+    ]
+    if box_format in ("centre", "unnamed"):
+        nodes += [
+            graph_node("Add", "corner_sum", ["x1y1", "x2y2"], ["sum"]),
+            graph_node("Mul", "corner_mean", ["sum", "half"], ["cxy"]),
+            graph_node("Sub", "corner_span", ["x2y2", "x1y1"], ["wh"]),
+            graph_node("Concat", "box_concat", ["cxy", "wh"], ["box"]),
+        ]
+    elif box_format == "corner":
+        nodes.append(graph_node("Concat", "box_concat", ["x1y1", "x2y2"], ["box"]))
+    else:
+        nodes.append(graph_node("Concat", "box_concat", ["x1y1", "x1y1"], ["box"]))
+
+    box = "box"
+    if unnamed:
+        box = "box_flat"
+        nodes.append(graph_node("Reshape", "box_reshape", ["shape", "box"], [box]))
+
+    nodes.append(onnx_node("Mul", PPU_BBOX_NODE, [box, "strides"], ["bbox_out"]))
+    graph = b"".join(proto_bytes(1, node) for node in nodes)
+    graph += proto_bytes(5, b"weights")
+    return (
+        proto_number(1, 10)  # ir_version
+        + proto_bytes(2, b"onnx_frontend_compiler")  # producer_name
+        + proto_bytes(7, graph)
+    )
+
+
+def write_dxnn(
+    directory, ppu, layers, name="model.dxnn", table=True, box_format=None
+) -> str:
+    """A minimal .dxnn as DX-RT's parsers read it: the container header, a
+    compile_config carrying `ppu`, and either the PPU tensor table with one
+    entry per (layer, anchor) as a v8 file has, or with `table` False only
+    the rmap_info listing of the PPU output tensors, as a v7 file has.
+    `layers` is (grid_w, grid_h, entries) per scale, finest first; an
+    anchor-free scale has one entry, and grid_h 1 means a flattened
+    (1, cells, channels) tensor. `box_format`, "centre" or "corner", adds
+    the compiled graph that says how the head writes its boxes, along with
+    the compile_config layer naming the node the PPU reads them from."""
+    if box_format is not None and "layer" not in ppu:
+        ppu = dict(ppu, layer=[{"bbox": PPU_BBOX_NODE, "cls_conf": "/head/Sigmoid"}])
+
+    compile_config = json.dumps({"compile_version": "2.4.0", "ppu": ppu}).encode()
+    graph = onnx_box_graph(box_format) if box_format is not None else b""
+
+    if table:
+        part = bytearray(struct.pack("<BBBB", 1, sum(n for _, _, n in layers), 0, 0))
+        for conv, (grid_w, grid_h, entries) in enumerate(layers):
+            for anchor in range(entries):
+                part += struct.pack(
+                    "<HHfBBBBBBBB",
+                    128,
+                    80,
+                    0.001,
+                    conv,
+                    anchor,
+                    0,
+                    1,
+                    0,
+                    grid_w,
+                    grid_h,
+                    0,
+                )
+    else:
+        outputs = []
+        for conv, (grid_w, grid_h, entries) in enumerate(layers):
+            for anchor in range(entries):
+                name_ = f"PPU_Transpose_Output_{conv}"
+                if entries > 1:
+                    name_ += f"_anchor_{anchor}"
+                shape = [1, grid_w, 127] if grid_h == 1 else [1, grid_h, grid_w, 128]
+                outputs.append({"name": name_, "shape": shape, "layout": "PPU_YOLO"})
+        part = json.dumps({"inputs": [], "outputs": outputs}).encode()
+
+    data = {
+        "compile_config": {
+            "type": "str",
+            "offset": 0,
+            "size": len(compile_config),
+        },
+        "compiled_data": {
+            "M1A_4K": {
+                "npu_0": {
+                    "rmap": {"type": "bytes", "offset": 0, "size": 0},
+                    "ppu" if table else "rmap_info": {
+                        "type": "bytes" if table else "str",
+                        "offset": len(compile_config),
+                        "size": len(part),
+                    },
+                }
+            }
+        },
+    }
+    if graph:
+        data["vis_npu_models"] = {
+            "npu_0": {
+                "type": "bytes",
+                "offset": len(compile_config) + len(part),
+                "size": len(graph),
+            }
+        }
+
+    index = json.dumps(
+        {
+            "version": 8 if table else 7,
+            "signature": "DXNN",
+            "size": 8192,
+            "data": data,
+        }
+    ).encode()
+
+    path = os.path.join(directory, name)
+    with open(path, "wb") as model:
+        model.write(b"DXNN" + struct.pack("<I", 8))
+        model.write(index.ljust(8192 - 8, b"\0"))
+        model.write(compile_config + part + graph)
+
+    return path
+
+
+def layout_of(shapes, num_classes, ppu=False, dynamic_output=False) -> YoloLayout:
+    return infer_yolo_layout(shapes, num_classes, ppu, dynamic_output).layout
+
+
+class TestDeepxModelFile(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        self.addCleanup(self.tmp.cleanup)
+
+    def layout(self, ppu, layers, **kwargs) -> PpuLayout | None:
+        return read_ppu_layout(write_dxnn(self.tmp.name, ppu, layers, **kwargs))
+
+    def test_the_head_kind_and_one_grid_per_scale_are_read(self):
+        cases = {
+            "anchor-based: three anchors per scale collapse to one grid each": (
+                (ANCHOR_BASED_PPU, THREE_SCALE_ANCHORS, {}),
+                PpuLayout(anchor_based=True, grids=((80, 80), (40, 40), (20, 20))),
+            ),
+            "anchor-free, one scale flattened into a single tensor": (
+                (ANCHOR_FREE_PPU, ONE_SCALE_FREE, {"box_format": "centre"}),
+                PpuLayout(anchor_based=False, grids=((100, 84),), centre_boxes=True),
+            ),
+            "a head kind the compiler does not name still yields the scales": (
+                ({"type": 7}, [(80, 80, 3), (40, 40, 3)], {}),
+                PpuLayout(anchor_based=None, grids=((80, 80), (40, 40))),
+            ),
+            "no table: the per-anchor split says anchor-based": (
+                ({"num_classes": 80}, THREE_SCALE_ANCHORS, {"table": False}),
+                PpuLayout(anchor_based=True, grids=((80, 80), (40, 40), (20, 20))),
+            ),
+            "no table: compile_config places the scales ahead of the names": (
+                (
+                    {
+                        "type": 1,
+                        "outputs": {
+                            f"PPU_Transpose_Output_{i}": {"conv_idx": 2 - i}
+                            for i in range(3)
+                        },
+                    },
+                    [(20, 20, 1), (40, 40, 1), (80, 80, 1)],
+                    {"table": False},
+                ),
+                PpuLayout(anchor_based=False, grids=((80, 80), (40, 40), (20, 20))),
+            ),
+            "no table: every scale in one (1, cells, channels) tensor": (
+                (ANCHOR_FREE_PPU, [(8400, 1, 1)], {"table": False}),
+                PpuLayout(anchor_based=False, grids=((8400, 1),)),
+            ),
+        }
+
+        for head, ((ppu, layers, kwargs), expected) in cases.items():
+            with self.subTest(head=head):
+                self.assertEqual(self.layout(ppu, layers, **kwargs), expected)
+
+    def test_the_box_format_comes_from_the_compiled_graph(self):
+        for box_format, centre in (
+            ("centre", True),
+            ("corner", False),
+            ("unnamed", True),
+        ):
+            with self.subTest(box_format=box_format):
+                self.assertEqual(
+                    self.layout(ANCHOR_FREE_PPU, ONE_SCALE_FREE, box_format=box_format),
+                    PpuLayout(
+                        anchor_based=False, grids=((100, 84),), centre_boxes=centre
+                    ),
+                )
+
+    def test_a_box_format_the_graph_does_not_answer_is_left_open(self):
+        cases = {
+            "no compiled graph at all": (ANCHOR_FREE_PPU, ONE_SCALE_FREE, {}),
+            "a graph without the node compile_config names": (
+                dict(ANCHOR_FREE_PPU, layer=[{"bbox": "/head/Missing"}]),
+                ONE_SCALE_FREE,
+                {"box_format": "centre"},
+            ),
+            "a graph that builds its boxes from neither shape": (
+                ANCHOR_FREE_PPU,
+                ONE_SCALE_FREE,
+                {"box_format": "neither"},
+            ),
+            "a graph section the reader cannot make sense of": (
+                ANCHOR_FREE_PPU,
+                ONE_SCALE_FREE,
+                {"box_format": "broken"},
+            ),
+            "a grid-decoded head, where the question does not arise": (
+                ANCHOR_FREE_PPU,
+                THREE_SCALE_FREE,
+                {"box_format": "centre"},
+            ),
+        }
+
+        for graph, (ppu, layers, kwargs) in cases.items():
+            with self.subTest(graph=graph):
+                self.assertIsNone(self.layout(ppu, layers, **kwargs).centre_boxes)
+
+    def test_nothing_is_read_from_a_model_without_ppu_metadata(self):
+        self.assertIsNone(read_ppu_layout(os.path.join(self.tmp.name, "missing")))
+
+        path = write_dxnn(self.tmp.name, None, [(80, 80, 3)])
+        self.assertIsNone(read_ppu_layout(path))
+
+        with open(path, "wb") as model:
+            model.write(b"ONNX" + b"\0" * 100)
+        self.assertIsNone(read_ppu_layout(path))
+
+        path = write_dxnn(self.tmp.name, ANCHOR_BASED_PPU, [(80, 80, 3), (40, 40, 3)])
+        os.truncate(path, os.path.getsize(path) - 40)
+        self.assertIsNone(read_ppu_layout(path))
+
+
+class TestDeepxLayoutInference(unittest.TestCase):
+    def test_a_shape_and_class_count_pick_one_layout(self):
+        cases = {
+            "anchor-free, four columns ahead of the classes": (
+                [(1, 84, 8400)],
+                80,
+                YoloLayout.anchor_free,
+                84,
+            ),
+            "anchor-free, row-major": ([(1, 8400, 84)], 80, YoloLayout.anchor_free, 84),
+            "anchor-based, an objectness column as well": (
+                [(1, 25200, 85)],
+                80,
+                YoloLayout.anchor,
+                85,
+            ),
+            "anchor-based, channel-major": (
+                [(1, 85, 25200)],
+                80,
+                YoloLayout.anchor,
+                85,
+            ),
+            "NMS in the head, a fixed run of corner records": (
+                [(1, 300, 6)],
+                80,
+                YoloLayout.nms_in_head,
+                None,
+            ),
+            # only the label map can tell these two apart
+            "85 columns with 81 classes is anchor-free": (
+                [(1, 8400, 85)],
+                81,
+                YoloLayout.anchor_free,
+                85,
+            ),
+            "85 columns with 80 classes is anchor-based": (
+                [(1, 8400, 85)],
+                80,
+                YoloLayout.anchor,
+                85,
+            ),
+            # 6 columns is all three layouts, told apart by the row count
+            "6 columns and thousands of rows, one class": (
+                [(1, 25200, 6)],
+                1,
+                YoloLayout.anchor,
+                6,
+            ),
+            "6 columns and thousands of rows, two classes": (
+                [(1, 8400, 6)],
+                2,
+                YoloLayout.anchor_free,
+                6,
+            ),
+            "6 columns and a few hundred rows, one class": (
+                [(1, 300, 6)],
+                1,
+                YoloLayout.nms_in_head,
+                None,
+            ),
+            "7 columns with two classes is only anchor-based": (
+                [(1, 8400, 7)],
+                2,
+                YoloLayout.anchor,
+                7,
+            ),
+            # a square output matches the same width on both axes, which must
+            # not read as two candidate layouts
+            "a square output is not ambiguous with itself": (
+                [(1, 85, 85)],
+                80,
+                YoloLayout.anchor,
+                85,
+            ),
+            "three NCHW maps with 255 channels are feature maps": (
+                [(1, 255, 80, 80), (1, 255, 40, 40), (1, 255, 20, 20)],
+                80,
+                YoloLayout.multipart,
+                None,
+            ),
+        }
+
+        for head, (shapes, num_classes, layout, columns) in cases.items():
+            with self.subTest(head=head):
+                output = infer_yolo_layout(shapes, num_classes, False, False)
+
+                self.assertIs(output.layout, layout)
+                if columns is not None:
+                    self.assertEqual(output.columns, columns)
+
+    def test_the_runtime_flags_outrank_the_shapes(self):
+        self.assertIs(layout_of([(8400,)], 80, ppu=True), YoloLayout.ppu)
+        self.assertIs(
+            layout_of([(1, -1, 6)], 80, dynamic_output=True), YoloLayout.nms_in_head
+        )
+
+    def test_a_shape_no_layout_fits_is_refused_with_the_reason(self):
+        cases = {
+            "fits two layouts": ([(1, 84, 6)], 80),
+            "labelmap_path": ([(1, 8400, 84)], 91),
+            "no output tensor": ([], 80),
+            "255 channels": (
+                [(1, 80, 80, 255), (1, 40, 40, 255), (1, 20, 20, 255)],
+                80,
+            ),
+            "feature maps": ([(1, 8400, 80), (1, 8400, 4)], 80),
+            "80-class": ([(1, 24, 80, 80), (1, 24, 40, 40), (1, 24, 20, 20)], 3),
+        }
+
+        for reason, (shapes, num_classes) in cases.items():
+            with (
+                self.subTest(reason=reason),
+                self.assertRaisesRegex(ValueError, reason),
+            ):
+                layout_of(shapes, num_classes)
+
+
+class TestDeepxOutputValidation(unittest.TestCase):
+    def test_the_raw_yolox_head_is_read_for_the_configured_input(self):
+        for shapes, size in (
+            ([(1, 8400, 85)], 640),
+            ([(1, 85, 8400)], 640),
+            # 52*52 + 26*26 + 13*13 cells at 416
+            ([(1, 3549, 85)], 416),
+        ):
+            with self.subTest(shapes=shapes, size=size):
+                self.assertEqual(validate_yolox_outputs(shapes, 80, size, size), 85)
+
+    def test_another_head_under_yolox_is_refused_with_the_reason(self):
+        cases = {
+            "width and height": ([(1, 8400, 85)], 80, 416),
+            "labelmap_path": ([(1, 8400, 85)], 91, 640),
+            "yolo-generic": ([(1, 8400, 80), (1, 8400, 4)], 80, 640),
+        }
+
+        for reason, (shapes, num_classes, size) in cases.items():
+            with (
+                self.subTest(reason=reason),
+                self.assertRaisesRegex(ValueError, reason),
+            ):
+                validate_yolox_outputs(shapes, num_classes, size, size)
+
+    def test_the_label_map_bounds_the_class_count(self):
+        self.assertEqual(class_count({0: "person", 79: "toothbrush"}), 80)
+
+        with self.assertRaisesRegex(ValueError, "labelmap_path"):
+            class_count({})
+
+
+class TestDeepxRawDecode(unittest.TestCase):
+    def anchor_rows(self, rows) -> list:
+        out = np.zeros((1, len(rows), 85), dtype=np.float32)
+
+        for i, (cx, cy, w, h, obj, label, score) in enumerate(rows):
+            out[0, i, 0:4] = [cx, cy, w, h]
+            out[0, i, 4] = obj
+            out[0, i, 5 + label] = score
+
+        return [out]
+
+    def test_an_anchor_based_head_becomes_normalized_corners(self):
+        outputs = self.anchor_rows([(320.0, 160.0, 64.0, 32.0, 0.8, 3, 0.5)])
+
+        detections = decode_raw_anchor(outputs, 640, 640, 0.25, 0.45)
+
+        self.assertEqual(detections[0][0], 3)
+        self.assertAlmostEqual(detections[0][1], 0.4, places=5)
+        self.assertAlmostEqual(detections[0][2], 144 / 640, places=5)
+        self.assertAlmostEqual(detections[0][3], 288 / 640, places=5)
+        self.assertAlmostEqual(detections[0][4], 176 / 640, places=5)
+        self.assertAlmostEqual(detections[0][5], 352 / 640, places=5)
+
+    def test_a_channel_major_export_is_read_by_column_count(self):
+        outputs = self.anchor_rows([(320.0, 160.0, 64.0, 32.0, 1.0, 3, 0.9)])
+
+        detections = decode_raw_anchor(
+            [np.swapaxes(outputs[0], 1, 2)], 640, 640, 0.25, 0.45, columns=85
+        )
+
+        self.assertEqual(detections[0][0], 3)
+        self.assertAlmostEqual(detections[0][3], 288 / 640, places=5)
+
+    def test_rows_below_the_combined_threshold_are_dropped(self):
+        # 0.4 * 0.5 = 0.2, under the threshold both parts clear on their own
+        outputs = self.anchor_rows([(320.0, 320.0, 40.0, 80.0, 0.4, 3, 0.5)])
+
+        self.assertTrue(np.all(decode_raw_anchor(outputs, 640, 640, 0.25, 0.45) == 0))
+
+    def test_at_most_twenty_detections_are_returned(self):
+        rows = [(20.0 + 24 * i, 320.0, 16.0, 16.0, 1.0, i % 80, 0.9) for i in range(25)]
+
+        detections = decode_raw_anchor(self.anchor_rows(rows), 640, 640, 0.25, 0.45)
+
+        self.assertEqual(detections.shape, (20, 6))
+        self.assertEqual(int((detections[:, 1] > 0).sum()), 20)
+
+    def test_an_nms_in_head_output_is_read_without_running_nms(self):
+        out = np.array(
+            [
+                [
+                    [100.0, 100.0, 200.0, 200.0, 0.9, 2.0],
+                    [102.0, 102.0, 202.0, 202.0, 0.8, 2.0],
+                    [300.0, 300.0, 400.0, 400.0, 0.1, 5.0],
+                ]
+            ],
+            dtype=np.float32,
+        )
+
+        detections = decode_raw_nms_in_head([out], 640, 640, 0.25)
+
+        self.assertEqual(detections[0][0], 2)
+        self.assertAlmostEqual(detections[0][1], 0.9, places=5)
+        self.assertAlmostEqual(detections[0][3], 100 / 640, places=5)
+        self.assertEqual(detections[1][0], 2)
+        self.assertAlmostEqual(detections[1][1], 0.8, places=5)
+        self.assertTrue(np.all(detections[2] == 0))
+
+        empty = np.zeros((1, 0, 6), dtype=np.float32)
+        self.assertTrue(np.all(decode_raw_nms_in_head([empty], 640, 640, 0.25) == 0))
+
+
+class TestDeepxConfig(unittest.TestCase):
+    def test_a_device_string_resolves_to_an_npu_index(self):
+        for configured, index in (("PCIe:1", 1), ("2", 2), ("", 0)):
+            with self.subTest(device=configured):
+                self.assertEqual(resolve_device(configured), index)
+
+    def test_a_device_that_is_not_an_index_is_rejected(self):
+        """The whole string has to be an index. Reading only the tail would
+        take the 1 out of "PCIe:0,PCIe:1" and bind to an NPU the config never
+        named, and several NPUs are configured as separate devices entries."""
+        for configured in (
+            "PCIe:the-fast-one",
+            "PCIe:0,PCIe:1",
+            "0,1",
+            "PCIe:0 PCIe:1",
+            "PCIe:-1",
+            "PCIe:",
+        ):
+            with self.subTest(device=configured):
+                with self.assertRaises(ValueError):
+                    resolve_device(configured)
+
+                with self.assertRaises(ValidationError):
+                    DeepxDetectorConfig(type="deepx", device=configured)
+
+    def test_a_bad_device_is_refused_where_the_config_is_parsed(self):
+        """parse_device builds the detector config to surface a bad device at
+        startup, so the comma-separated form fails there rather than binding a
+        detector process to the wrong NPU."""
+        self.assertEqual(parse_device("deepx:PCIe:1").device, "PCIe:1")
+
+        for raw in ("deepx:PCIe:0,PCIe:1", "deepx:the-fast-one"):
+            with self.subTest(raw=raw), self.assertRaises(DeviceParseError):
+                parse_device(raw)
+
+    def test_a_device_string_builds_this_detector_config(self):
+        """Frigate turns a `deepx:PCIe:0` entry into the detector config with
+        the model already attached, which is the path app.py takes; a bare
+        constructor call does not exercise it."""
+        config = build_detector_config(
+            parse_device("deepx:PCIe:0"), model_with_type(ModelTypeEnum.yologeneric)
+        )
+
+        self.assertIsInstance(config, DeepxDetectorConfig)
+        self.assertEqual(config.device, "PCIe:0")
+        self.assertEqual(config.model.model_type, ModelTypeEnum.yologeneric)
+
+    def test_the_runtime_manifest_pins_its_wheels_to_the_version(self):
+        """A PyPI path carries a per-file digest, so bumping DXRT_VERSION has
+        to rewrite the whole URL; a stale one installs the old wheel and fails
+        the sha256 on every user's first start."""
+        self.assertEqual(DEEPX_MANIFEST.version, DXRT_VERSION)
+
+        for artifact in DEEPX_MANIFEST.artifacts:
+            with self.subTest(url=artifact.url):
+                self.assertIn(f"dx_engine-{DXRT_VERSION}-", artifact.url)
+
+    def test_a_model_less_config_still_validates(self):
+        config = DeepxDetectorConfig(type="deepx")
+
+        self.assertIsNone(config.model)
+
+
+class DeepxDetectorTestCase(unittest.TestCase):
+    def detector(
+        self,
+        model_type=ModelTypeEnum.yologeneric,
+        outputs_info=None,
+        ppu=False,
+        dynamic=False,
+        model_path="/nonexistent/model.dxnn",
+    ) -> DeepxDetector:
+        dx_engine = MagicMock()
+        dx_engine.Configuration.ITEM.SERVICE = object()
+        session = dx_engine.InferenceEngine.return_value
+        session.get_output_tensors_info.return_value = (
+            [{"shape": [8400]}] if ppu else outputs_info or []
+        )
+        session.is_ppu.return_value = ppu
+        session.has_dynamic_output.return_value = dynamic
+
+        config = DeepxDetectorConfig(type="deepx")
+        config.model = model_with_type(model_type)
+        config.model.path = model_path
+
+        with (
+            # the detector writes the endpoint into the environment, which the
+            # rest of the suite shares when it runs in one process
+            patch.dict(os.environ),
+            patch.dict(sys.modules, {"dx_engine": dx_engine}),
+            patch.object(DeepxDetector, "activate_dependencies"),
+            patch("os.path.isfile", return_value=True),
+        ):
+            return DeepxDetector(config)
+
+    def ppu_detector(
+        self, ppu, layers, model_type=ModelTypeEnum.yologeneric, box_format=None
+    ) -> DeepxDetector:
+        tmp = tempfile.TemporaryDirectory()
+        self.addCleanup(tmp.cleanup)
+
+        return self.detector(
+            model_type,
+            ppu=True,
+            model_path=write_dxnn(tmp.name, ppu, layers, box_format=box_format),
+        )
+
+    def detect(self, detector, outputs) -> np.ndarray:
+        detector.session.run.return_value = outputs
+        return detector.detect_raw(np.zeros((1, 640, 640, 3), np.uint8))
+
+
+class TestDeepxModelType(DeepxDetectorTestCase):
+    def test_a_model_type_with_no_decoder_is_rejected(self):
+        for model_type in (ModelTypeEnum.ssd, ModelTypeEnum.dfine):
+            with (
+                self.subTest(model_type=model_type),
+                self.assertRaisesRegex(ValueError, model_type.value),
+            ):
+                self.detector(model_type)
+
+    def test_supported_model_types_are_accepted(self):
+        for model_type, outputs_info in (
+            (ModelTypeEnum.yologeneric, [{"shape": [1, 84, 8400]}]),
+            (ModelTypeEnum.yolox, [{"shape": [1, 8400, 85]}]),
+        ):
+            with self.subTest(model_type=model_type):
+                detector = self.detector(model_type, outputs_info)
+
+                self.assertEqual(detector.model_type, model_type)
+
+
+class TestDeepxDetectorLoad(DeepxDetectorTestCase):
+    def test_the_layout_is_settled_at_load(self):
+        detector = self.detector(ModelTypeEnum.yologeneric, [{"shape": [1, 84, 8400]}])
+        self.assertIs(detector.output.layout, YoloLayout.anchor_free)
+        self.assertEqual(detector.output.columns, 84)
+
+        detector = self.detector(ModelTypeEnum.yolox, [{"shape": [1, 8400, 85]}])
+        self.assertIs(detector.output.layout, YoloLayout.yolox)
+
+        for model_type in (ModelTypeEnum.yologeneric, ModelTypeEnum.yolox):
+            with self.subTest(model_type=model_type):
+                detector = self.ppu_detector(
+                    ANCHOR_FREE_PPU, THREE_SCALE_FREE, model_type=model_type
+                )
+
+                self.assertIs(detector.output.layout, YoloLayout.ppu)
+                self.assertEqual(detector.ppu_layout.scale_count, 3)
+
+    def test_a_model_the_detector_cannot_decode_is_refused_at_load(self):
+        cases = {
+            "Cannot decode DEEPX model": lambda: self.detector(
+                ModelTypeEnum.yologeneric, [{"shape": [1, 8400, 7]}]
+            ),
+            "DX-COM 2.4.0": lambda: self.detector(ModelTypeEnum.yologeneric, ppu=True),
+            "face and pose": lambda: self.ppu_detector({"type": 2}, [(80, 80, 1)]),
+            "centre and size or as two corners": lambda: self.ppu_detector(
+                ANCHOR_FREE_PPU, ONE_SCALE_FREE
+            ),
+        }
+
+        for reason, load in cases.items():
+            with (
+                self.subTest(reason=reason),
+                self.assertRaisesRegex(ValueError, reason),
+            ):
+                load()
+
+
+class TestDeepxDetectRaw(DeepxDetectorTestCase):
+    def test_a_ppu_record_decodes_by_the_head_in_the_model(self):
+        cases = {
+            "anchor-based, layer 2 of 3 at stride 32, the 373x326 anchor": (
+                (ANCHOR_BASED_PPU, THREE_SCALE_ANCHORS, None),
+                build_ppu_record((0.6, 0.4, 0.3, 0.7), label=5),
+                (5, (0.0, 0.380094, 0.864187, 0.589906)),
+            ),
+            "anchor-based, layer 0 of 3 at stride 8, the 16x30 anchor": (
+                (ANCHOR_BASED_PPU, THREE_SCALE_ANCHORS, None),
+                build_ppu_record((0.6, 0.4, 0.3, 0.7), grid=(7, 9, 1, 0)),
+                (0, (None, 0.116750, None, None)),
+            ),
+            "anchor-based, two scales: layer 0 is stride 16, not stride 8": (
+                (ANCHOR_BASED_PPU, TWO_SCALE_ANCHORS, None),
+                build_ppu_record((0.6, 0.4, 0.3, 0.7), grid=(7, 9, 1, 0)),
+                (0, (0.141156, 0.236031, 0.223844, 0.248969)),
+            ),
+            "anchor-free, three scales: cell (10, 9) of stride 32": (
+                (ANCHOR_FREE_PPU, THREE_SCALE_FREE, None),
+                build_ppu_record((1.2, 0.5, 1.0, 0.5), grid=(9, 10, 0, 2), label=7),
+                (7, (0.433782, 0.492043, 0.516218, 0.627957)),
+            ),
+            "anchor-free, one scale: the box fields are already pixels": (
+                (ANCHOR_FREE_PPU, ONE_SCALE_FREE, "centre"),
+                build_ppu_record((320.0, 160.0, 64.0, 32.0), label=3),
+                (3, (144 / 640, 288 / 640, 176 / 640, 352 / 640)),
+            ),
+            "anchor-free, one scale: a sub-pixel box stays sub-pixel": (
+                (ANCHOR_FREE_PPU, ONE_SCALE_FREE, "centre"),
+                build_ppu_record((0.6, 0.4, 0.3, 0.7)),
+                (0, (None, 0.45 / 640, None, None)),
+            ),
+            "a corner-format head reads the record as two corners": (
+                (ANCHOR_FREE_PPU, ONE_SCALE_FREE, "corner"),
+                build_ppu_record((100.0, 50.0, 300.0, 250.0), label=2),
+                (2, (50 / 640, 100 / 640, 250 / 640, 300 / 640)),
+            ),
+            "a centre-format head reads it as a centre and size": (
+                (ANCHOR_FREE_PPU, ONE_SCALE_FREE, "centre"),
+                build_ppu_record((100.0, 50.0, 300.0, 250.0), label=2),
+                (2, (0.0, 0.0, 175 / 640, 250 / 640)),
+            ),
+        }
+
+        for head, ((ppu, layers, box_format), record, (label, box)) in cases.items():
+            with self.subTest(head=head):
+                detector = self.ppu_detector(ppu, layers, box_format=box_format)
+
+                detections = self.detect(detector, [record])
+
+                self.assertEqual(detections[0][0], label)
+                for i, expected in enumerate(box, start=2):
+                    if expected is not None:
+                        self.assertAlmostEqual(detections[0][i], expected, places=5)
+
+    def test_the_strides_come_from_the_grids_in_the_model(self):
+        detector = self.ppu_detector(
+            ANCHOR_FREE_PPU, [(40, 40, 1), (20, 20, 1), (10, 10, 1)]
+        )
+
+        detections = self.detect(
+            detector,
+            [build_ppu_record((1.2, 0.5, 1.0, 0.5), grid=(9, 10, 0, 0), label=7)],
+        )
+
+        # layer 0 is stride 16: centre (11.2, 9.5) * 16, size e * 16 x sqrt(e) * 16
+        self.assertEqual(detections[0][0], 7)
+        self.assertAlmostEqual(detections[0][2], (152 - np.exp(0.5) * 8) / 640, 5)
+        self.assertAlmostEqual(detections[0][3], (179.2 - np.exp(1.0) * 8) / 640, 5)
+
+    def test_the_head_stays_what_the_model_said_across_frames(self):
+        detector = self.ppu_detector(ANCHOR_BASED_PPU, THREE_SCALE_ANCHORS)
+
+        first = self.detect(
+            detector, [build_ppu_record((0.6, 0.4, 0.3, 0.7), grid=(7, 9, 1, 0))]
+        )
+        # layer 0 of 3: stride 8, anchor 16x30
+        self.assertAlmostEqual(first[0][3], 0.116750, places=5)
+
+        second = self.detect(
+            detector, [build_ppu_record((0.5, 0.5, 0.4, 0.4), grid=(3, 4, 0, 1))]
+        )
+        # layer 1 of 3: stride 16, anchor 30x61, not layer 1 of 2
+        self.assertAlmostEqual(second[0][3], 0.0975, places=5)
+
+        detector = self.ppu_detector(
+            ANCHOR_FREE_PPU, ONE_SCALE_FREE, box_format="centre"
+        )
+        record = [build_ppu_record((100.0, 50.0, 300.0, 250.0), label=2)]
+        # centre (100, 50), size 300 x 250: the right edge lands at 250
+        for frame in range(2):
+            with self.subTest(frame=frame):
+                self.assertAlmostEqual(
+                    self.detect(detector, record)[0][5], 250 / 640, places=5
+                )
+
+    def test_records_the_head_cannot_place_come_back_empty(self):
+        cases = {
+            "a level or box the anchor table does not carry": (
+                THREE_SCALE_ANCHORS,
+                [build_ppu_record((0.6, 0.4, 0.3, 0.7), grid=(7, 9, 2, 5))],
+            ),
+            "a scale count Frigate has no anchor table for": (
+                FOUR_SCALE_ANCHORS,
+                [build_ppu_record((0.6, 0.4, 0.3, 0.7))],
+            ),
+            "a record below the score threshold": (
+                THREE_SCALE_ANCHORS,
+                [build_ppu_record((0.6, 0.4, 0.3, 0.7), score=0.1)],
+            ),
+            "an unexpected record width": (
+                THREE_SCALE_ANCHORS,
+                [np.zeros((1, 3, 16), dtype=np.uint8)],
+            ),
+            **{
+                f"no records at all, shaped {shape}": (
+                    THREE_SCALE_ANCHORS,
+                    [np.zeros(shape, dtype=np.uint8)],
+                )
+                for shape in ((1, 0, PPU_RECORD_SIZE), (0, PPU_RECORD_SIZE), (0,))
+            },
+        }
+
+        for output, (layers, outputs) in cases.items():
+            with self.subTest(output=output):
+                detector = self.ppu_detector(ANCHOR_BASED_PPU, layers)
+
+                self.assertTrue(np.all(self.detect(detector, outputs) == 0))
+
+    def test_a_yolox_raw_head_is_decoded_through_the_grid(self):
+        detector = self.detector(ModelTypeEnum.yolox, [{"shape": [1, 8400, 85]}])
+
+        # cell (x 10, y 5) of the stride-8 grid is row 5 * 80 + 10
+        tensor = np.zeros((1, 8400, 85), np.float32)
+        tensor[0, 410, :5] = [0.5, 0.5, np.log(4.0), np.log(2.0), 0.9]
+        tensor[0, 410, 5 + 7] = 0.8
+
+        detections = self.detect(detector, [tensor])
+
+        # centre (84, 44), size 32 x 16
+        self.assertEqual(detections[0][0], 7)
+        self.assertAlmostEqual(detections[0][1], 0.72, places=5)
+        self.assertAlmostEqual(detections[0][2], 36 / 640, places=5)
+        self.assertAlmostEqual(detections[0][3], 68 / 640, places=5)
+        self.assertAlmostEqual(detections[0][4], 52 / 640, places=5)
+        self.assertAlmostEqual(detections[0][5], 100 / 640, places=5)
+
+        detector = self.detector(ModelTypeEnum.yolox, [{"shape": [1, 85, 8400]}])
+        detections = self.detect(detector, [np.swapaxes(tensor, 1, 2)])
+        self.assertAlmostEqual(detections[0][5], 100 / 640, places=5)
+
+    def test_a_raw_head_comes_back_as_frigates_detection_rows(self):
+        detector = self.detector(ModelTypeEnum.yologeneric, [{"shape": [1, 84, 8400]}])
+        output = np.zeros((1, 84, 8400), dtype=np.float32)
+        output[0, 0:4, 0] = [320.0, 160.0, 64.0, 32.0]
+        output[0, 4 + 2, 0] = 0.9
+
+        detections = self.detect(detector, [output])
+
+        self.assertEqual(detections.shape, (20, 6))
+        self.assertEqual(detections[0][0], 2)
+        self.assertAlmostEqual(detections[0][1], 0.9, places=5)
+        self.assertAlmostEqual(detections[0][3], 288 / 640, places=5)

frigate/test/test_detector_hardware.py

@@ -184,6 +184,26 @@ class TestAccelerators(HardwareProbeTestCase):
         )
         self.assertFalse(memryx.unlimited)
 
+    def test_each_deepx_node_is_a_unit(self):
+        write(os.path.join(self.dev_root, "dxrt0"))
+        write(os.path.join(self.dev_root, "dxrt1"))
+
+        deepx = self.probe()["deepx"]
+
+        self.assertEqual(
+            [unit.device for unit in deepx.units],
+            ["deepx:PCIe:0", "deepx:PCIe:1"],
+        )
+
+    def test_a_deepx_npu_is_unlimited(self):
+        # the host daemon multiplexes, so one module takes several processes
+        write(os.path.join(self.dev_root, "dxrt0"))
+
+        self.assertTrue(self.probe()["deepx"].unlimited)
+
+    def test_no_deepx_is_reported_without_a_node(self):
+        self.assertNotIn("deepx", self.probe())
+
     def test_a_supported_rockchip_soc_is_reported(self):
         write(
             os.path.join(self.proc_root, "device-tree", "compatible"),

frigate/test/test_dst_transitions.py

@@ -0,0 +1,54 @@
+"""Tests for get_dst_transitions."""
+
+import datetime
+import unittest
+
+from frigate.util.time import get_dst_transitions
+
+
+class TestDstTransitions(unittest.TestCase):
+    def test_dst_transition_splits_periods_at_the_transition(self):
+        start = datetime.datetime(2026, 3, 7, 12, tzinfo=datetime.UTC).timestamp()
+        end = start + 2 * 86400
+        spring = datetime.datetime(2026, 3, 8, 7, tzinfo=datetime.UTC).timestamp()
+        self.assertEqual(
+            get_dst_transitions("America/New_York", start, end),
+            [(start, spring, -18000), (spring, end, -14400)],
+        )
+
+    def test_dst_transition_is_not_reported_a_day_late(self):
+        # local midnight on the day of the change used to report the
+        # transition a full day after it actually happened
+        start = datetime.datetime(2024, 3, 10, 5, tzinfo=datetime.UTC).timestamp()
+        end = start + 3 * 86400
+        spring = datetime.datetime(2024, 3, 10, 7, tzinfo=datetime.UTC).timestamp()
+        self.assertEqual(
+            get_dst_transitions("America/New_York", start, end),
+            [(start, spring, -18000), (spring, end, -14400)],
+        )
+
+    def test_dst_transition_after_the_last_daily_probe_is_found(self):
+        start = datetime.datetime(2024, 11, 2, 12, tzinfo=datetime.UTC).timestamp()
+        end = datetime.datetime(2024, 11, 3, 10, tzinfo=datetime.UTC).timestamp()
+        fall = datetime.datetime(2024, 11, 3, 6, tzinfo=datetime.UTC).timestamp()
+        self.assertEqual(
+            get_dst_transitions("America/New_York", start, end),
+            [(start, fall, -14400), (fall, end, -18000)],
+        )
+
+    def test_no_transition_returns_a_single_period(self):
+        start = datetime.datetime(2026, 6, 1, tzinfo=datetime.UTC).timestamp()
+        end = start + 5 * 86400
+        self.assertEqual(
+            get_dst_transitions("America/New_York", start, end),
+            [(start, end, -14400)],
+        )
+
+    def test_invalid_zone_retains_utc_fallback(self):
+        self.assertEqual(
+            get_dst_transitions("Invalid/Timezone", 100, 200), [(100, 200, 0)]
+        )
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)

frigate/test/test_genai_providers.py

@@ -360,9 +360,70 @@ class TestOllamaProvider(unittest.TestCase):
         from frigate.genai.plugins.ollama import _normalize_multimodal_content
 
         text, images = _normalize_multimodal_content(MULTIMODAL_MESSAGES[-1]["content"])
-        self.assertIn("live image", text)
-        self.assertEqual(len(images), 1)
-        self.assertEqual(images[0], b"\xff\xd8\xff\xd9")
+        self.assertEqual(
+            text, "Here is the current live image from camera 'front'.\n[img]"
+        )
+        self.assertEqual(images, [b"\xff\xd8\xff\xd9"])
+
+    def test_normalize_keeps_text_and_image_order(self):
+        from frigate.genai.plugins.ollama import _normalize_multimodal_content
+
+        text, images = _normalize_multimodal_content(
+            [
+                {"type": "text", "text": "intro"},
+                {"type": "text", "text": "Frame 1"},
+                {"type": "image_url", "image_url": {"url": _IMAGE_DATA_URI}},
+                {"type": "text", "text": "Frame 2"},
+                {"type": "image_url", "image_url": {"url": _IMAGE_DATA_URI}},
+            ]
+        )
+        self.assertEqual(text, "intro\nFrame 1\n[img]\nFrame 2\n[img]")
+        self.assertEqual(len(images), 2)
+
+    def test_send_uses_chat_with_captions_before_each_image(self):
+        client = self._client()
+        client.provider = MagicMock()
+        client.provider.chat.return_value = {
+            "message": {"content": '{"ok": true}'},
+            "done": True,
+            "done_reason": "stop",
+        }
+        client._supports_thinking_cache = False
+
+        result = client._send(
+            "prompt",
+            [b"a", b"b"],
+            {"type": "json_schema", "json_schema": {"schema": {"type": "object"}}},
+            image_captions=["Frame 1 of 2", "Frame 2 of 2"],
+        )
+
+        self.assertEqual(result, '{"ok": true}')
+        client.provider.generate.assert_not_called()
+        params = client.provider.chat.call_args.kwargs
+        self.assertEqual(
+            params["messages"],
+            [
+                {
+                    "role": "user",
+                    "content": "prompt\nFrame 1 of 2\n[img]\nFrame 2 of 2\n[img]",
+                    "images": [b"a", b"b"],
+                }
+            ],
+        )
+        self.assertEqual(params["format"], {"type": "object"})
+        self.assertNotIn("think", params)
+
+    def test_send_without_captions_puts_images_after_prompt(self):
+        client = self._client()
+        client.provider = MagicMock()
+        client.provider.chat.return_value = {"message": {"content": "ok"}, "done": True}
+        client._supports_thinking_cache = False
+
+        client._send("prompt", [b"a"])
+
+        message = client.provider.chat.call_args.kwargs["messages"][0]
+        self.assertEqual(message["content"], "prompt\n[img]")
+        self.assertEqual(message["images"], [b"a"])
 
 
 # ---------------------------------------------------------------------------
@@ -519,6 +580,281 @@ class TestLlamaCppProvider(unittest.TestCase):
         client = self._validated_client(4096, {"context_size": 32768})
         self.assertEqual(client.get_context_size(), 32768)
 
+    def test_list_models_dedupes_alias_matching_id(self):
+        client = self._client()
+        models_data = [
+            {"id": "qwen3-asr", "aliases": ["qwen3-asr"]},
+            {"id": "gemma", "aliases": ["gemma", "g4"]},
+        ]
+        with patch.object(client, "_fetch_models_data", return_value=models_data):
+            self.assertEqual(client.list_models(), ["g4", "gemma", "qwen3-asr"])
+
+
+# ---------------------------------------------------------------------------
+# transcribe role
+# ---------------------------------------------------------------------------
+WAV_BYTES = b"RIFF$\x00\x00\x00WAVEfmt "
+
+
+class TestOpenAITranscribe(unittest.TestCase):
+    def _client(self):
+        return _make_client(
+            "openai",
+            model="gpt-4o-transcribe",
+            api_key="k",
+            base_url="http://localhost:9999/v1",
+            runtime_options={"temperature": 0.7},
+        )
+
+    def test_supports_transcription(self):
+        self.assertTrue(self._client().supports_transcription)
+
+    def test_passes_file_tuple_and_language(self):
+        client = self._client()
+        create = MagicMock(return_value="  hello there  ")
+        client.provider = SimpleNamespace(
+            audio=SimpleNamespace(transcriptions=SimpleNamespace(create=create))
+        )
+
+        self.assertEqual(client.transcribe(WAV_BYTES, language="en"), "hello there")
+
+        kwargs = create.call_args.kwargs
+        self.assertEqual(kwargs["model"], "gpt-4o-transcribe")
+        self.assertEqual(kwargs["file"], ("audio.wav", WAV_BYTES, "audio/wav"))
+        self.assertEqual(kwargs["language"], "en")
+        self.assertEqual(kwargs["response_format"], "text")
+
+    def test_does_not_forward_runtime_options(self):
+        """runtime_options are chat parameters; /audio/transcriptions rejects them."""
+        client = self._client()
+        create = MagicMock(return_value="hi")
+        client.provider = SimpleNamespace(
+            audio=SimpleNamespace(transcriptions=SimpleNamespace(create=create))
+        )
+
+        client.transcribe(WAV_BYTES)
+
+        self.assertNotIn("temperature", create.call_args.kwargs)
+
+    def test_gpt_transcribe_uses_languages_array(self):
+        """gpt-transcribe replaced `language` with a `languages` array."""
+        client = _make_client("openai", model="gpt-transcribe", api_key="k")
+        create = MagicMock(return_value="hi")
+        client.provider = SimpleNamespace(
+            audio=SimpleNamespace(transcriptions=SimpleNamespace(create=create))
+        )
+
+        client.transcribe(WAV_BYTES, language="en")
+
+        kwargs = create.call_args.kwargs
+        self.assertEqual(kwargs["extra_body"], {"languages": ["en"]})
+        # sending both fields is rejected by the API
+        self.assertNotIn("language", kwargs)
+
+    def test_older_models_use_singular_language(self):
+        for model in ("gpt-4o-transcribe", "whisper-1"):
+            with self.subTest(model):
+                client = _make_client("openai", model=model, api_key="k")
+                create = MagicMock(return_value="hi")
+                client.provider = SimpleNamespace(
+                    audio=SimpleNamespace(transcriptions=SimpleNamespace(create=create))
+                )
+
+                client.transcribe(WAV_BYTES, language="en")
+
+                kwargs = create.call_args.kwargs
+                self.assertEqual(kwargs["language"], "en")
+                self.assertNotIn("extra_body", kwargs)
+
+    def test_object_response_form(self):
+        client = self._client()
+        create = MagicMock(return_value=SimpleNamespace(text="hi"))
+        client.provider = SimpleNamespace(
+            audio=SimpleNamespace(transcriptions=SimpleNamespace(create=create))
+        )
+
+        self.assertEqual(client.transcribe(WAV_BYTES), "hi")
+
+    def test_error_returns_none(self):
+        client = self._client()
+        create = MagicMock(side_effect=RuntimeError("boom"))
+        client.provider = SimpleNamespace(
+            audio=SimpleNamespace(transcriptions=SimpleNamespace(create=create))
+        )
+
+        self.assertIsNone(client.transcribe(WAV_BYTES))
+
+
+class TestAzureOpenAITranscribe(unittest.TestCase):
+    def _client(self):
+        return _make_client(
+            "azure_openai",
+            model="my-deployment",
+            api_key="k",
+            base_url="https://example.openai.azure.com/?api-version=2024-06-01",
+        )
+
+    def test_routes_through_azure_client(self):
+        from openai import AzureOpenAI
+
+        client = self._client()
+        self.assertIsInstance(client.provider, AzureOpenAI)
+        self.assertTrue(client.supports_transcription)
+
+    def test_transcribe_inherited(self):
+        client = self._client()
+        create = MagicMock(return_value="azure text")
+        client.provider = SimpleNamespace(
+            audio=SimpleNamespace(transcriptions=SimpleNamespace(create=create))
+        )
+
+        self.assertEqual(client.transcribe(WAV_BYTES, language="fr"), "azure text")
+        self.assertEqual(create.call_args.kwargs["model"], "my-deployment")
+
+
+class TestGeminiTranscribe(unittest.TestCase):
+    def _client(self):
+        return _make_client("gemini", model="gemini-2.0-flash", api_key="k")
+
+    def test_supports_transcription(self):
+        self.assertTrue(self._client().supports_transcription)
+
+    def test_sends_audio_part(self):
+        client = self._client()
+        generate = MagicMock(return_value=SimpleNamespace(text=" spoken words "))
+        client.provider = SimpleNamespace(
+            models=SimpleNamespace(generate_content=generate)
+        )
+
+        self.assertEqual(client.transcribe(WAV_BYTES, language="en"), "spoken words")
+
+        contents = generate.call_args.kwargs["contents"]
+        audio_parts = [
+            p for p in contents if getattr(p, "inline_data", None) is not None
+        ]
+        self.assertEqual(len(audio_parts), 1)
+        self.assertEqual(audio_parts[0].inline_data.mime_type, "audio/wav")
+        self.assertEqual(audio_parts[0].inline_data.data, WAV_BYTES)
+
+    def test_oversized_payload_is_skipped(self):
+        from frigate.genai.plugins.gemini import GEMINI_MAX_INLINE_BYTES
+
+        client = self._client()
+        generate = MagicMock()
+        client.provider = SimpleNamespace(
+            models=SimpleNamespace(generate_content=generate)
+        )
+
+        self.assertIsNone(client.transcribe(b"\x00" * (GEMINI_MAX_INLINE_BYTES + 1)))
+        generate.assert_not_called()
+
+
+class TestLlamaCppTranscribe(unittest.TestCase):
+    def _client(self, supports_audio: bool):
+        cfg = GenAIConfig(
+            provider="llamacpp",
+            model="m",
+            base_url="http://localhost:9999",
+        )
+        info = {
+            "context_size": 4096,
+            "supports_vision": False,
+            "supports_audio": supports_audio,
+            "supports_tools": False,
+            "supports_reasoning": False,
+            "media_marker": "<__media__>",
+        }
+        cls = PROVIDERS[GenAIProviderEnum.llamacpp]
+        with patch.object(cls, "_get_model_info", return_value=info):
+            return cls(cfg, timeout=5)
+
+    def test_supports_transcription_tracks_supports_audio(self):
+        self.assertTrue(self._client(True).supports_transcription)
+        self.assertFalse(self._client(False).supports_transcription)
+
+    @staticmethod
+    def _transcriptions_response(text: str = " transcript "):
+        response = MagicMock()
+        response.status_code = 200
+        response.json.return_value = {"text": text}
+        return response
+
+    @staticmethod
+    def _chat_response(content: str = " fallback transcript "):
+        response = MagicMock()
+        response.status_code = 200
+        response.json.return_value = {"choices": [{"message": {"content": content}}]}
+        return response
+
+    def test_posts_multipart_to_transcriptions(self):
+        client = self._client(True)
+
+        with patch.object(
+            client, "_post", return_value=self._transcriptions_response()
+        ) as post:
+            self.assertEqual(client.transcribe(WAV_BYTES, language="en"), "transcript")
+
+        self.assertTrue(post.call_args.args[0].endswith("/v1/audio/transcriptions"))
+        self.assertEqual(
+            post.call_args.kwargs["files"]["file"],
+            ("audio.wav", WAV_BYTES, "audio/wav"),
+        )
+        self.assertEqual(post.call_args.kwargs["data"]["language"], "en")
+
+    def test_omits_language_when_not_set(self):
+        """An unset language is what lets the model detect one itself."""
+        client = self._client(True)
+
+        with patch.object(
+            client, "_post", return_value=self._transcriptions_response()
+        ) as post:
+            client.transcribe(WAV_BYTES)
+
+        self.assertNotIn("language", post.call_args.kwargs["data"])
+
+    def test_falls_back_to_chat_completions_on_404(self):
+        """Servers predating llama.cpp#21863 have no transcriptions route."""
+        client = self._client(True)
+        missing = MagicMock()
+        missing.status_code = 404
+
+        with patch.object(
+            client, "_post", side_effect=[missing, self._chat_response()]
+        ) as post:
+            self.assertEqual(
+                client.transcribe(WAV_BYTES, language="en"), "fallback transcript"
+            )
+
+        urls = [call.args[0] for call in post.call_args_list]
+        self.assertTrue(urls[0].endswith("/v1/audio/transcriptions"))
+        self.assertTrue(urls[1].endswith("/v1/chat/completions"))
+
+        payload = post.call_args_list[1].kwargs["json"]
+        content = payload["messages"][0]["content"]
+        audio_parts = [p for p in content if p["type"] == "input_audio"]
+        self.assertEqual(len(audio_parts), 1)
+        self.assertEqual(audio_parts[0]["input_audio"]["format"], "wav")
+        self.assertEqual(
+            base64.b64decode(audio_parts[0]["input_audio"]["data"]), WAV_BYTES
+        )
+
+    def test_audio_unsupported_returns_none(self):
+        client = self._client(False)
+
+        with patch.object(client, "_post") as post:
+            self.assertIsNone(client.transcribe(WAV_BYTES))
+
+        post.assert_not_called()
+
+
+class TestBaseClientTranscribe(unittest.TestCase):
+    """Providers that don't implement the role must be inert, not broken."""
+
+    def test_ollama_reports_and_returns_nothing(self):
+        client = _make_client("ollama", model="llava", base_url="http://localhost:9999")
+        self.assertFalse(client.supports_transcription)
+        self.assertIsNone(client.transcribe(WAV_BYTES, language="en"))
+
 
 if __name__ == "__main__":
     unittest.main()

frigate/test/test_motion_detector.py

@@ -73,6 +73,35 @@ class TestImprovedMotionDetector(unittest.TestCase):
             "Motion boxes should be empty when scene change exceeds skip threshold",
         )
 
+    def _bright_frame(self, offset: int) -> np.ndarray:
+        """Produce a bright frame with a small dark object that moves."""
+        frame = np.full((self.frame_shape[0], self.frame_shape[1]), 200, dtype=np.uint8)
+        x = 10 + (offset * 5) % 60
+        frame[40:60, x : x + 15] = 20
+        return frame
+
+    def test_skip_motion_threshold_recovers_after_skip(self):
+        """Skipped frames must still be blended into the background.
+
+        A bright scene differs from the zeroed background across the whole
+        frame, so the first frames are skipped. The background has to catch up
+        anyway, otherwise motion detection never returns.
+        """
+        self.config.skip_motion_threshold = 0.5
+        self.config.improve_contrast = False
+        self.detector.config = self.config
+        self.detector.update_mask()
+
+        boxes = [len(self.detector.detect(self._bright_frame(i))) for i in range(40)]
+
+        self.assertEqual(boxes[0], 0, "First frame should exceed the skip threshold")
+        self.assertGreater(
+            self.detector.avg_frame.max(),
+            0,
+            "Background was never updated while frames were skipped",
+        )
+        self.assertTrue(any(boxes), "Motion detection never recovered after a skip")
+
     def test_skip_motion_threshold_does_not_affect_calibration(self):
         """Even when skipping, the detector should go into calibrating state."""
         self.config.skip_motion_threshold = 0.4

frigate/test/test_ptz_onvif.py

@@ -15,6 +15,8 @@ Also covers the inverse direction: the ptz movement timestamps must not be writt
 for a camera that has autotracking off, because nothing clears them back out.
 """
 
+import asyncio
+import threading
 import unittest
 from unittest.mock import AsyncMock, MagicMock
 
@@ -234,5 +236,40 @@ class TestManualRelativeMoveMetrics(unittest.IsolatedAsyncioTestCase):
         )
 
 
+class TestOnvifClose(unittest.TestCase):
+    """close() must release everything on the loop, since whatever it leaves is
+    garbage collected during interpreter shutdown, where the resulting warnings
+    fail to log and fill the shutdown output with logging errors."""
+
+    def setUp(self) -> None:
+        self.controller = _make_controller(autotracking_enabled=False)
+        self.onvif = self.controller.cams[CAMERA]["onvif"]
+        self.onvif.close = AsyncMock()
+        self.controller.config_subscriber = MagicMock()
+        self.controller.loop = asyncio.new_event_loop()
+        self.controller.loop_thread = threading.Thread(
+            target=self.controller._run_event_loop, daemon=True
+        )
+        self.controller.loop_thread.start()
+        self.addCleanup(self.controller.loop.close)
+
+    def test_close_closes_camera_sessions(self) -> None:
+        self.controller.close()
+
+        self.onvif.close.assert_awaited_once()
+
+    def test_close_cancels_tasks_left_on_the_loop(self) -> None:
+        async def forever() -> None:
+            while True:
+                await asyncio.sleep(1)
+
+        poll = asyncio.run_coroutine_threadsafe(forever(), self.controller.loop)
+
+        self.controller.close()
+
+        self.assertTrue(poll.cancelled())
+        self.assertFalse(self.controller.loop_thread.is_alive())
+
+
 if __name__ == "__main__":
     unittest.main()

frigate/test/test_restart_frigate.py

@@ -0,0 +1,55 @@
+"""Tests for restarting frigate under s6."""
+
+import signal
+import unittest
+from unittest.mock import MagicMock, patch
+
+import psutil
+
+from frigate.util.services import restart_frigate
+
+
+class TestRestartFrigate(unittest.TestCase):
+    def _s6_process(self) -> MagicMock:
+        proc = MagicMock()
+        proc.name.return_value = "s6-svscan"
+        return proc
+
+    @patch("frigate.util.services.os.kill")
+    @patch("frigate.util.services.psutil.Process")
+    def test_terminates_s6_when_permitted(self, mock_process, mock_kill):
+        proc = self._s6_process()
+        mock_process.return_value = proc
+
+        restart_frigate()
+
+        proc.terminate.assert_called_once()
+        mock_kill.assert_not_called()
+
+    @patch("frigate.util.services.os.getpid", return_value=99)
+    @patch("frigate.util.services.os.kill")
+    @patch("frigate.util.services.psutil.Process")
+    def test_exits_self_when_s6_signal_is_denied(
+        self, mock_process, mock_kill, _mock_getpid
+    ):
+        """Running unprivileged, frigate cannot signal root's s6-svscan."""
+        proc = self._s6_process()
+        proc.terminate.side_effect = psutil.AccessDenied(pid=1, name="s6-svscan")
+        mock_process.return_value = proc
+
+        restart_frigate()
+
+        mock_kill.assert_called_once_with(99, signal.SIGINT)
+
+    @patch("frigate.util.services.os.getpid", return_value=99)
+    @patch("frigate.util.services.os.kill")
+    @patch("frigate.util.services.psutil.Process")
+    def test_exits_self_without_s6(self, mock_process, mock_kill, _mock_getpid):
+        proc = MagicMock()
+        proc.name.return_value = "init"
+        mock_process.return_value = proc
+
+        restart_frigate()
+
+        proc.terminate.assert_not_called()
+        mock_kill.assert_called_once_with(99, signal.SIGINT)

frigate/test/test_review_annotations.py

@@ -0,0 +1,316 @@
+"""Tests for tracker-derived review frame annotations."""
+
+import unittest
+
+from frigate.data_processing.post.review_annotations import (
+    annotations_by_frame,
+    build_timeline,
+    describe_heading,
+    describe_position,
+    event_name,
+    path_legs,
+    path_moments,
+)
+
+
+def straight_path(
+    start: tuple[float, float],
+    end: tuple[float, float],
+    steps: int,
+    t0: float,
+) -> list:
+    """A path_data-shaped trajectory travelling in a straight line."""
+    return [
+        [
+            [
+                start[0] + (end[0] - start[0]) * i / (steps - 1),
+                start[1] + (end[1] - start[1]) * i / (steps - 1),
+            ],
+            t0 + i,
+        ]
+        for i in range(steps)
+    ]
+
+
+class TestDescribers(unittest.TestCase):
+    def test_describes_frame_corners(self):
+        self.assertEqual(describe_position(0.9, 0.9), "the bottom right of the frame")
+        self.assertEqual(describe_position(0.1, 0.1), "the top left of the frame")
+        self.assertEqual(describe_position(0.5, 0.5), "the middle of the frame")
+
+    def test_heading_treats_falling_y_as_up(self):
+        self.assertEqual(describe_heading(0.0, -0.5), "up")
+        self.assertEqual(describe_heading(0.0, 0.5), "down")
+        self.assertEqual(describe_heading(-0.5, 0.0), "left")
+
+    def test_heading_combines_axes(self):
+        self.assertEqual(describe_heading(-0.5, -0.5), "up and left")
+
+
+class TestPathLegs(unittest.TestCase):
+    def test_straight_travel_is_one_leg(self):
+        points = [(0.9 - 0.05 * i, 0.6, float(i)) for i in range(10)]
+        self.assertEqual(len(path_legs(points)), 1)
+
+    def test_out_and_back_is_two_legs(self):
+        out = [(0.9 - 0.05 * i, 0.6, float(i)) for i in range(8)]
+        back = [(0.55 + 0.05 * i, 0.6, 8.0 + i) for i in range(8)]
+        self.assertEqual(len(path_legs(out + back)), 2)
+
+    def test_jitter_in_place_produces_no_legs(self):
+        # A subject standing still wobbles by a couple of percent; without the
+        # minimum leg distance this became a burst of contradictory turns.
+        points = [(0.5 + 0.01 * (i % 2), 0.5, float(i)) for i in range(20)]
+        self.assertEqual(path_legs(points), [])
+
+    def test_single_point_path_has_no_moments(self):
+        self.assertEqual(path_moments([[[0.5, 0.5], 1.0]]), [])
+
+    def test_malformed_path_is_ignored(self):
+        self.assertEqual(path_moments([[0.5, 1.0], [0.6, 2.0]]), [])
+
+
+class TestEventNames(unittest.TestCase):
+    def test_unnamed_objects_use_an_indefinite_article(self):
+        self.assertEqual(event_name({"label": "person"}), "a person")
+        self.assertEqual(event_name({"label": "animal"}), "an animal")
+
+    def test_sub_labeled_objects_use_their_name(self):
+        event = {"label": "waste_bin", "sub_label": "Compost"}
+        self.assertEqual(event_name(event), 'waste bin "Compost"')
+
+    def test_repeated_objects_are_never_numbered(self):
+        # Whether these are the same subject is unknown, so the notes must not
+        # imply either answer.
+        events = [
+            {
+                "id": f"1789481994.68448{i}-abcdef",
+                "label": "person",
+                "sub_label": None,
+                "start_time": float(i * 10),
+                "end_time": float(i * 10 + 50),
+                "zones": [],
+                "path_data": straight_path((0.9, 0.6), (0.4, 0.3), 10, i * 10.0),
+            }
+            for i in range(3)
+        ]
+        phrases = [p for _, p in build_timeline(events, span_end=100.0)]
+        self.assertTrue(all(p.startswith("a person ") for p in phrases))
+        self.assertFalse(any("#" in p for p in phrases))
+
+
+class TestTimeline(unittest.TestCase):
+    def setUp(self):
+        self.event = {
+            "id": "1789481994.684479-lpyc2z",
+            "label": "person",
+            "sub_label": None,
+            "start_time": 0.0,
+            "end_time": 20.0,
+            "zones": ["front_yard"],
+            "path_data": straight_path((0.9, 0.6), (0.4, 0.3), 10, 1.0),
+        }
+
+    def test_timeline_is_ordered_and_bounded(self):
+        timeline = build_timeline([self.event], span_end=30.0)
+        times = [t for t, _ in timeline]
+        self.assertEqual(times, sorted(times))
+        self.assertTrue(all(t <= 30.0 for t in times))
+
+    def test_track_identifiers_never_appear(self):
+        timeline = build_timeline([self.event], span_end=30.0)
+        joined = " ".join(phrase for _, phrase in timeline)
+        self.assertNotIn("track", joined.lower())
+        self.assertNotIn(self.event["id"], joined)
+
+    def test_object_still_tracked_at_end_gets_no_closing_note(self):
+        # Its state at the end is visible in the last frame, so nothing is said.
+        timeline = build_timeline([self.event], span_end=15.0)
+        self.assertEqual(
+            [p for _, p in timeline],
+            ["a person first detected at the right of the frame, moving up and left"],
+        )
+
+    def test_object_ending_inside_the_clip_is_no_longer_detected(self):
+        timeline = build_timeline([self.event], span_end=40.0)
+        phrases = [p for _, p in timeline]
+        self.assertTrue(any("no longer detected" in p for p in phrases))
+
+    def test_moments_past_the_last_frame_are_dropped(self):
+        # The subject keeps moving after the final sampled frame; those notes
+        # describe nothing the model can see.
+        late = dict(self.event)
+        out = straight_path((0.9, 0.6), (0.4, 0.6), 8, 100.0)
+        back = straight_path((0.4, 0.6), (0.9, 0.6), 8, 108.0)
+        late["path_data"] = out + back
+        late["start_time"] = 100.0
+        timeline = build_timeline([late], span_end=105.0)
+        self.assertTrue(any("moving left" in p for _, p in timeline))
+        self.assertFalse(any("turns around" in p for _, p in timeline))
+
+
+def track(event_id, label, start, path, sub_label=None, end=None):
+    return {
+        "id": event_id,
+        "label": label,
+        "sub_label": sub_label,
+        "start_time": start,
+        "end_time": end if end is not None else start + 500.0,
+        "zones": [],
+        "path_data": path,
+    }
+
+
+class TestArrival(unittest.TestCase):
+    def test_objects_arriving_together_keep_separate_notes(self):
+        person = track(
+            "1789481994.684479-lpyc2z",
+            "person",
+            0.0,
+            straight_path((0.9, 0.63), (0.58, 0.34), 10, 0.1),
+        )
+        bin_ = track(
+            "1789481995.063395-vlzd7q",
+            "waste_bin",
+            0.3,
+            straight_path((0.91, 0.64), (0.6, 0.33), 10, 0.4),
+            sub_label="Compost",
+        )
+        phrases = [p for _, p in build_timeline([person, bin_], 100.0)]
+        self.assertEqual(
+            phrases,
+            [
+                "a person first detected at the right of the frame, moving up and left",
+                'waste bin "Compost" first detected at the right of the frame, '
+                "moving up and left",
+            ],
+        )
+
+    def test_object_detected_well_before_it_moves_gets_separate_notes(self):
+        # path_data always keeps the first two samples, so a bin sitting in
+        # the yard opens its leg long before it is picked up.
+        path = [[[0.91, 0.64], 0.4]] + straight_path(
+            (0.91, 0.64), (0.6, 0.33), 10, 20.4
+        )
+        bin_ = track(
+            "1789481995.063395-vlzd7q", "waste_bin", 0.3, path, sub_label="Compost"
+        )
+        timeline = build_timeline([bin_], 100.0)
+        self.assertEqual(
+            timeline[0],
+            (0.3, 'waste bin "Compost" first detected at the right of the frame'),
+        )
+        self.assertAlmostEqual(timeline[1][0], 21.4)
+        self.assertEqual(
+            timeline[1][1],
+            'waste bin "Compost" starts moving up and left from the right of the frame',
+        )
+
+
+class TestStateChanges(unittest.TestCase):
+    def test_stationary_and_active_rows_become_notes(self):
+        bin_ = track(
+            "1789481995.063395-vlzd7q",
+            "waste_bin",
+            0.3,
+            straight_path((0.91, 0.64), (0.6, 0.33), 10, 0.4),
+            sub_label="Compost",
+        )
+        changes = [
+            {"timestamp": 20.0, "source_id": bin_["id"], "class_type": "stationary"},
+            {"timestamp": 30.0, "source_id": bin_["id"], "class_type": "active"},
+            {"timestamp": 35.0, "source_id": bin_["id"], "class_type": "entered_zone"},
+            {
+                "timestamp": 40.0,
+                "source_id": "someone-else",
+                "class_type": "stationary",
+            },
+        ]
+        timeline = build_timeline([bin_], 100.0, state_changes=changes)
+        self.assertEqual(
+            timeline[1:],
+            [
+                (20.0, 'waste bin "Compost" has stopped moving'),
+                (30.0, 'waste bin "Compost" starts moving again'),
+            ],
+        )
+
+    def test_state_changes_after_the_last_frame_are_dropped(self):
+        bin_ = track(
+            "1789481995.063395-vlzd7q",
+            "waste_bin",
+            0.3,
+            straight_path((0.91, 0.64), (0.6, 0.33), 10, 0.4),
+            sub_label="Compost",
+        )
+        changes = [
+            {"timestamp": 200.0, "source_id": bin_["id"], "class_type": "stationary"}
+        ]
+        timeline = build_timeline([bin_], 100.0, state_changes=changes)
+        self.assertFalse(any("stopped" in p for _, p in timeline))
+
+
+class TestNoAssumedState(unittest.TestCase):
+    def test_no_note_for_where_movement_ends(self):
+        # Ending a leftward walk still in the right third was read as a turn
+        # back to the right, and "stops moving" was read as standing still.
+        moments = path_moments(straight_path((0.95, 0.6), (0.7, 0.4), 6, 0.0))
+        self.assertEqual(
+            [p for _, p in moments],
+            ["starts moving up and left from the right of the frame"],
+        )
+
+    def test_notes_never_claim_an_object_is_stationary(self):
+        # A track still open at the last frame says nothing about motion.
+        event = {
+            "id": "1789482056.695307-3uhf47",
+            "label": "person",
+            "sub_label": None,
+            "start_time": 0.0,
+            "end_time": 500.0,
+            "zones": ["front_yard"],
+            "path_data": straight_path((0.9, 0.6), (0.4, 0.3), 10, 1.0),
+        }
+        joined = " ".join(p for _, p in build_timeline([event], span_end=20.0))
+        self.assertNotIn("stationary", joined)
+        self.assertNotIn("stops", joined)
+        self.assertNotIn("leaves", joined)
+        self.assertNotIn("still", joined)
+
+
+class TestLateEvents(unittest.TestCase):
+    def test_objects_first_detected_after_the_last_frame_are_skipped(self):
+        # Without this the arrival lands on the final frame, which was
+        # captured before the object appeared.
+        late = track(
+            "1789481999.000000-latear",
+            "person",
+            50.0,
+            straight_path((0.9, 0.6), (0.4, 0.3), 10, 50.1),
+        )
+        self.assertEqual(build_timeline([late], span_end=40.0), [])
+        self.assertEqual(
+            annotations_by_frame(build_timeline([late], 40.0), [0.0, 40.0]), {}
+        )
+
+
+class TestFrameBucketing(unittest.TestCase):
+    def test_moment_attaches_to_the_following_frame(self):
+        frame_times = [0.0, 10.0, 20.0, 30.0]
+        buckets = annotations_by_frame([(12.0, "something happened")], frame_times)
+        self.assertEqual(buckets, {2: ["something happened"]})
+
+    def test_moment_on_a_frame_boundary_uses_that_frame(self):
+        buckets = annotations_by_frame([(10.0, "x")], [0.0, 10.0, 20.0])
+        self.assertEqual(buckets, {1: ["x"]})
+
+    def test_moment_after_the_last_frame_falls_on_the_last_frame(self):
+        buckets = annotations_by_frame([(99.0, "x")], [0.0, 10.0])
+        self.assertEqual(buckets, {1: ["x"]})
+
+    def test_no_frames_yields_no_buckets(self):
+        self.assertEqual(annotations_by_frame([(1.0, "x")], []), {})
+
+
+if __name__ == "__main__":
+    unittest.main()

frigate/track/norfair_tracker.py

@@ -324,9 +324,7 @@ class NorfairTracker(ObjectTracker):
         ):
             tracker = self.get_tracker(obj["label"])
             tracker.tracked_objects = [
-                o
-                for o in tracker.tracked_objects
-                if str(o.global_id) != track_id and o.hit_counter < 0
+                o for o in tracker.tracked_objects if str(o.global_id) != track_id
             ]
 
         del self.track_id_map[track_id]

frigate/util/audio.py

@@ -1,16 +1,61 @@
 """Utilities for creating and manipulating audio."""
 
+import io
 import logging
 import os
+import re
+import string
+import struct
 import subprocess as sp
+import wave
 
+import numpy as np
 from pathvalidate import sanitize_filename
 
-from frigate.const import CACHE_DIR, STREAM_TYPE_MAIN, STREAM_TYPE_SUB
+from frigate.const import (
+    AUDIO_SAMPLE_RATE,
+    CACHE_DIR,
+    STREAM_TYPE_MAIN,
+    STREAM_TYPE_SUB,
+)
 from frigate.models import Recordings
 
 logger = logging.getLogger(__name__)
 
+# Ceiling on the run of words the stitcher will treat as an overlap between two
+# consecutive windows. This is an audio-duration bound, not a linguistic one: a
+# window holds GENAI_WINDOW_CHUNKS * AUDIO_DURATION seconds of speech, so at a
+# fast talker's pace it tops out around this many words, and a whole window can
+# legitimately be redundant. The vendored whisper_streaming HypothesisBuffer
+# caps at 5, but there the n-gram is only a tie-break on top of word-level
+# timestamps; here it is the entire alignment, so 5 truncates real overlaps.
+# Sentinel meaning "let the model work out the language". The vendored
+# whisper_streaming code already uses this spelling, so it is the established
+# convention for the audio_transcription.language field.
+AUTO_LANGUAGE = "auto"
+
+MAX_STITCH_NGRAM = 16
+
+# How many trailing committed words the stitcher may discard to find an
+# alignment. Those words came from the newest audio, which the next window
+# re-covers, so when the provider got one of them wrong it blocks every
+# alignment and the whole phrase duplicates. Set to 0 to make committed text
+# strictly append-only.
+MAX_STITCH_REVISE = 3
+
+# A revision deletes text that was already published, so it has to clear a
+# higher bar than a plain append: a single coincidentally shared word is not
+# enough evidence to throw committed words away.
+MIN_STITCH_REVISE_RUN = 2
+
+# ASR models often wrap their output in control markup. Qwen3-ASR, for example,
+# answers "language English<asr_text>Yeah, that works." A structural opening tag
+# marks where the transcript starts, so anything before the last one is metadata.
+# Closing tags (</x>) and pipe-delimited special tokens (<|endoftext|>) are
+# excluded: those mark where the text ends, so text before them must be kept.
+_OPENING_TAG = re.compile(r"<(?![/|])[^<>]*>")
+_ANY_TAG = re.compile(r"<[^<>]*>")
+
 
 def _get_recordings_for_range(
     camera_name: str, start_ts: float, end_ts: float, stream_type: str
@@ -117,7 +162,9 @@ def get_audio_from_recording(
             logger.debug(
                 f"Successfully extracted audio for {camera_name} from {start_ts} to {end_ts}"
             )
-            return process.stdout
+            # ffmpeg writes to a pipe, so it cannot seek back to patch the chunk
+            # sizes it reserved; repair them before any strict consumer sees them
+            return fix_wav_header(process.stdout)
         else:
             logger.error(f"Failed to extract audio: {process.stderr.decode()}")
             return None
@@ -129,3 +176,217 @@ def get_audio_from_recording(
             os.unlink(file_path)
         except OSError:
             pass
+
+
+def fix_wav_header(data: bytes) -> bytes:
+    """Recompute the RIFF and data chunk sizes in a WAV header.
+
+    ffmpeg writing to a non-seekable pipe cannot go back and patch the sizes it
+    reserved, so it leaves 0xFFFFFFFF placeholders. PyAV-based demuxers ignore
+    them, but strict validators may reject the file or read zero frames.
+
+    Args:
+        data: The complete WAV payload
+
+    Returns:
+        The payload with both sizes corrected, or unchanged if it is not a
+        parseable RIFF/WAVE stream
+    """
+    if len(data) < 12 or data[0:4] != b"RIFF" or data[8:12] != b"WAVE":
+        return data
+
+    out = bytearray(data)
+
+    # RIFF size covers everything after the 8-byte RIFF header
+    struct.pack_into("<I", out, 4, len(out) - 8)
+
+    # walk the chunk list to find "data"; every chunk is padded to even length
+    pos = 12
+
+    while pos + 8 <= len(out):
+        chunk_id = bytes(out[pos : pos + 4])
+        (chunk_size,) = struct.unpack_from("<I", out, pos + 4)
+
+        if chunk_id == b"data":
+            struct.pack_into("<I", out, pos + 4, len(out) - (pos + 8))
+            return bytes(out)
+
+        if chunk_size == 0xFFFFFFFF:
+            # an unpatched size before the data chunk leaves nothing to walk
+            break
+
+        pos += 8 + chunk_size + (chunk_size % 2)
+
+    return bytes(out)
+
+
+def pcm16_to_wav(samples: np.ndarray, sample_rate: int = AUDIO_SAMPLE_RATE) -> bytes:
+    """Wrap mono int16 PCM samples in a WAV container.
+
+    Args:
+        samples: The audio samples; converted to int16 if they are not already
+        sample_rate: Sample rate to declare in the header
+
+    Returns:
+        WAV bytes suitable for upload to a GenAI provider
+    """
+    if samples.dtype != np.int16:
+        samples = samples.astype(np.int16)
+
+    buffer = io.BytesIO()
+
+    with wave.open(buffer, "wb") as wav:
+        wav.setnchannels(1)
+        wav.setsampwidth(2)
+        wav.setframerate(sample_rate)
+        wav.writeframes(samples.tobytes())
+
+    return buffer.getvalue()
+
+
+def stitch_transcripts(committed: str, incoming: str) -> str:
+    """Append *incoming* to *committed*, dropping the speech they share.
+
+    Consecutive overlapped transcription windows re-transcribe the same audio at
+    their seam, so the tail of one and the newest one name the same words. Find
+    the longest run that is a suffix of *committed* and occurs anywhere in
+    *incoming*, then keep only what follows that run.
+
+    Searching all of *incoming* rather than just its start is what makes this
+    work in practice. The provider re-transcribes the shared audio independently
+    and often gets its first word or two different ("just gonna" one window,
+    "It's gonna" the next), which leaves the real overlap sitting in the middle
+    of *incoming*. A prefix-anchored match sees no overlap at all there and
+    duplicates the entire phrase.
+
+    Text-level rather than timestamp-level because only some providers return
+    word timings, and this has to work across all of them.
+
+    Args:
+        committed: The transcript accumulated so far
+        incoming: The newest window's transcript
+
+    Returns:
+        The combined transcript
+    """
+    incoming_words = incoming.split()
+
+    if not incoming_words:
+        return committed
+
+    committed_words = committed.split()
+
+    if not committed_words:
+        return " ".join(incoming_words)
+
+    committed_keys = [_overlap_key(word) for word in committed_words]
+    incoming_keys = [_overlap_key(word) for word in incoming_words]
+
+    length, consumed = _find_overlap(committed_keys, incoming_keys)
+
+    if length:
+        return " ".join(committed_words + incoming_words[consumed:])
+
+    # Nothing aligns. Retry against a shortened committed tail: a single word the
+    # provider got wrong at the end of the previous window otherwise blocks every
+    # alignment, and the entire re-transcribed phrase duplicates behind it.
+    best: tuple[int, int, int] | None = None
+
+    for drop in range(1, min(MAX_STITCH_REVISE, len(committed_keys) - 1) + 1):
+        length, consumed = _find_overlap(committed_keys[:-drop], incoming_keys)
+
+        if length < MIN_STITCH_REVISE_RUN:
+            continue
+
+        # longest run wins; ties go to the smallest revision
+        if best is None or length > best[0]:
+            best = (length, drop, consumed)
+
+    if best is None:
+        return " ".join(committed_words + incoming_words)
+
+    _, drop, consumed = best
+
+    return " ".join(committed_words[:-drop] + incoming_words[consumed:])
+
+
+def _find_overlap(
+    committed_keys: list[str], incoming_keys: list[str]
+) -> tuple[int, int]:
+    """Locate the speech *incoming* shares with the end of *committed*.
+
+    Returns the length of the longest run that is a suffix of *committed_keys*
+    and occurs anywhere in *incoming_keys*, along with the index just past that
+    run in *incoming_keys*. Returns ``(0, 0)`` when nothing matches.
+
+    Prefers the longest run so a real overlap is not cut short, and within one
+    length the earliest position, so a phrase genuinely spoken twice keeps its
+    second utterance.
+    """
+    max_run = min(MAX_STITCH_NGRAM, len(committed_keys), len(incoming_keys))
+
+    for length in range(max_run, 0, -1):
+        tail = committed_keys[-length:]
+
+        for start in range(len(incoming_keys) - length + 1):
+            if incoming_keys[start : start + length] == tail:
+                return length, start + length
+
+    return 0, 0
+
+
+def clean_transcript(text: str | None) -> str:
+    """Strip provider control markup and any preamble from a raw transcript.
+
+    A window with no speech often still comes back as the preamble alone
+    ("language English<asr_text>"), which must reduce to an empty string so
+    callers treat it as silence rather than committing it as spoken words.
+
+    Args:
+        text: The provider's raw response
+
+    Returns:
+        The transcript with markup removed and whitespace collapsed
+    """
+    if not text:
+        return ""
+
+    # everything up to and including the last opening tag is metadata
+    openings = list(_OPENING_TAG.finditer(text))
+
+    if openings:
+        text = text[openings[-1].end() :]
+
+    # drop closing tags and special tokens wherever they landed
+    text = _ANY_TAG.sub(" ", text)
+
+    return " ".join(text.split())
+
+
+def _overlap_key(word: str) -> str:
+    """Comparison key for overlap matching.
+
+    Providers re-transcribe the shared audio at a window seam independently, so
+    the same word routinely comes back capitalized differently or with different
+    edge punctuation ("work." vs "Work"). Those differences must not defeat the
+    match, but the original spelling is what gets kept in the output.
+    """
+    key = word.strip(string.punctuation).casefold()
+
+    # a token that is nothing but punctuation would otherwise match any other
+    return key or word
+
+
+def resolve_language(language: str | None) -> str | None:
+    """Turn a configured language into an explicit code, or None for auto-detect.
+
+    Args:
+        language: The configured value, possibly AUTO_LANGUAGE
+
+    Returns:
+        An ISO language code, or None when the backend should detect it
+    """
+    if not language or language == AUTO_LANGUAGE:
+        return None
+
+    return language

frigate/util/builtin.py

@@ -56,10 +56,13 @@ class EventsPerSecond:
             self._start = now
         # compute the (approximate) events in the last n seconds
         self.expire_timestamps(now)
-        seconds = min(now - self._start, self._last_n_seconds)
-        # avoid divide by zero
-        if seconds == 0:
-            seconds = 1
+        # rate over at least one second (or the whole window, if shorter),
+        # so a burst of events right after start() is not divided by a
+        # tiny window
+        seconds = max(
+            min(now - self._start, self._last_n_seconds),
+            min(1.0, self._last_n_seconds),
+        )
         return len(self._timestamps) / seconds
 
     # remove aged out timestamps

frigate/util/config.py

@@ -829,6 +829,57 @@ def rename_hailo_detector(
     return new_config
 
 
+def _camera_enables_transcription(camera: dict[str, Any]) -> bool:
+    """Whether a camera or one of its profiles turns audio transcription on."""
+    sections = [camera.get("audio_transcription")]
+    profiles = camera.get("profiles")
+
+    if isinstance(profiles, dict):
+        for profile in profiles.values():
+            if isinstance(profile, dict):
+                sections.append(profile.get("audio_transcription"))
+
+    return any(
+        isinstance(section, dict) and section.get("enabled") for section in sections
+    )
+
+
+def _migrate_transcription_language(config: dict[str, Any]) -> None:
+    """Pin English for configs written before the language default became auto.
+
+    audio_transcription.language used to default to "en", so a config that
+    turned transcription on without naming a language was transcribing English.
+    The default is now "auto" (let the model detect), which is better for new
+    users but would silently change behavior for existing ones, so write the old
+    value explicitly for anyone actually using the feature.
+    """
+    transcription = config.get("audio_transcription")
+
+    if isinstance(transcription, dict) and "language" in transcription:
+        # named a language already, so nothing was relying on the default
+        return
+
+    enabled = isinstance(transcription, dict) and bool(transcription.get("enabled"))
+
+    if not enabled:
+        enabled = any(
+            _camera_enables_transcription(camera)
+            for camera in config.get("cameras", {}).values()
+            if isinstance(camera, dict)
+        )
+
+    if not enabled:
+        return
+
+    if not isinstance(transcription, dict):
+        # a camera enabled it without a global section, which still picked up
+        # the global default
+        transcription = {}
+        config["audio_transcription"] = transcription
+
+    transcription["language"] = "en"
+
+
 def migrate_019_0(config: dict[str, dict[str, Any]]) -> dict[str, dict[str, Any]]:
     """Handle migrating Frigate config to 0.19-0."""
     new_config = rename_hailo_detector(config)
@@ -845,6 +896,8 @@ def migrate_019_0(config: dict[str, dict[str, Any]]) -> dict[str, dict[str, Any]
 
         new_config["cameras"][name] = camera_config
 
+    _migrate_transcription_language(new_config)
+
     new_config["version"] = "0.19-0"
     return new_config
 

frigate/util/ffmpeg.py

@@ -49,6 +49,9 @@ def start_or_restart_ffmpeg(
     if ffmpeg_process is not None:
         stop_ffmpeg(ffmpeg_process, logger)
 
+        # flush after the stop so the logs cover ffmpeg's output up to exit
+        logpipe.dump()
+
     if frame_size is None:
         process = sp.Popen(
             ffmpeg_cmd,

frigate/util/services.py

@@ -35,12 +35,19 @@ logger = logging.getLogger(__name__)
 
 def restart_frigate():
     proc = psutil.Process(1)
+
     # if this is running via s6, sigterm pid 1
     if proc.name() == "s6-svscan":
-        proc.terminate()
+        try:
+            proc.terminate()
+            return
+        except psutil.AccessDenied:
+            # frigate runs unprivileged, so it cannot signal root's s6-svscan.
+            # exiting this process instead runs frigate/finish, which halts s6
+            logger.debug("Not permitted to signal s6-svscan, exiting instead")
+
     # otherwise, just try and exit frigate
-    else:
-        os.kill(os.getpid(), signal.SIGINT)
+    os.kill(os.getpid(), signal.SIGINT)
 
 
 def print_stack(sig, frame):

frigate/util/time.py

@@ -2,6 +2,7 @@
 
 import datetime
 import logging
+import math
 from zoneinfo import ZoneInfoNotFoundError
 
 import pytz
@@ -43,9 +44,33 @@ def is_current_hour(timestamp: int) -> bool:
     return timestamp < start_of_next_hour
 
 
+def _utc_offset(tz: datetime.tzinfo, timestamp: float) -> float:
+    dt = datetime.datetime.fromtimestamp(timestamp, tz=datetime.UTC)
+    return dt.astimezone(tz).utcoffset().total_seconds()
+
+
+def _find_transition(
+    tz: datetime.tzinfo, lo: float, hi: float, lo_offset: float
+) -> float:
+    """Bisect (lo, hi] to the second where the UTC offset first differs from lo_offset."""
+    # whole seconds, so the midpoint always advances (a fractional bound can
+    # otherwise leave the midpoint sitting on lo) and lands on the transition
+    low = math.floor(lo)
+    high = math.ceil(hi)
+
+    while high - low > 1:
+        mid = (low + high) // 2
+        if _utc_offset(tz, mid) == lo_offset:
+            low = mid
+        else:
+            high = mid
+
+    return float(high)
+
+
 def get_dst_transitions(
     tz_name: str, start_time: float, end_time: float
-) -> list[tuple[float, float]]:
+) -> list[tuple[float, float, float]]:
     """
     Find DST transition points and return time periods with consistent offsets.
 
@@ -66,28 +91,24 @@ def get_dst_transitions(
 
     periods = []
     current = start_time
-
-    # Get initial offset
-    dt = datetime.datetime.utcfromtimestamp(current).replace(tzinfo=pytz.UTC)
-    local_dt = dt.astimezone(tz)
-    prev_offset = local_dt.utcoffset().total_seconds()
     period_start = start_time
+    prev_offset = _utc_offset(tz, current)
 
-    # Check each day for offset changes
-    while current <= end_time:
-        dt = datetime.datetime.utcfromtimestamp(current).replace(tzinfo=pytz.UTC)
-        local_dt = dt.astimezone(tz)
-        current_offset = local_dt.utcoffset().total_seconds()
+    # Probe at most a day ahead, capped at end_time so a transition after the
+    # last full day is still seen instead of silently kept in the last period.
+    while current < end_time:
+        next_probe = min(current + 86400, end_time)
+        next_offset = _utc_offset(tz, next_probe)
 
-        if current_offset != prev_offset:
-            # Found a transition - close previous period
-            periods.append((period_start, current, prev_offset))
-            period_start = current
-            prev_offset = current_offset
+        if next_offset != prev_offset:
+            transition = _find_transition(tz, current, next_probe, prev_offset)
+            periods.append((period_start, transition, prev_offset))
+            period_start = transition
+            prev_offset = _utc_offset(tz, transition)
+            current = transition
+        else:
+            current = next_probe
 
-        current += 86400  # Check daily
-
-    # Add final period
     periods.append((period_start, end_time, prev_offset))
 
     return periods

frigate/video/ffmpeg.py

@@ -319,9 +319,6 @@ class CameraWatchdog(threading.Thread):
                     f"Capture thread for {self.config.name} did not exit in time"
                 )
 
-        self.logger.error(
-            "The following ffmpeg logs include the last 100 lines prior to exit."
-        )
         self.logpipe.dump()
         self.logger.info("Restarting ffmpeg...")
         self.start_ffmpeg_detect()
@@ -543,7 +540,6 @@ class CameraWatchdog(threading.Thread):
                         f"{self.config.name}/status/{role.value}", "offline"
                     )
 
-                p["logpipe"].dump()
                 p["process"] = start_or_restart_ffmpeg(
                     p["cmd"], self.logger, p["logpipe"], ffmpeg_process=p["process"]
                 )

web/e2e/helpers/monaco.ts

@@ -56,3 +56,14 @@ export async function waitForErrorMarker(
     .poll(() => hasErrorMarkers(page), { timeout: timeoutMs })
     .toBe(true);
 }
+
+/**
+ * Append a trailing comment so the editor content differs from the
+ * loaded config. The save buttons are disabled until then.
+ */
+export async function makeMonacoEdit(page: Page): Promise<void> {
+  const editor = page.locator(".monaco-editor").first();
+  await editor.click();
+  await page.keyboard.press("ControlOrMeta+End");
+  await page.keyboard.type("# edited by e2e", { delay: 0 });
+}

web/e2e/specs/config-editor.spec.ts

@@ -11,6 +11,7 @@ import { installWsFrameCapture, waitForWsFrame } from "../helpers/ws-frames";
 import { grantClipboardPermissions, readClipboard } from "../helpers/clipboard";
 import {
   getMonacoVisibleText,
+  makeMonacoEdit,
   replaceMonacoValue,
   waitForErrorMarker,
 } from "../helpers/monaco";
@@ -71,6 +72,7 @@ test.describe("Config Editor — Save @medium", () => {
     await expect(frigateApp.page.locator(".monaco-editor").first()).toBeVisible(
       { timeout: 15_000 },
     );
+    await makeMonacoEdit(frigateApp.page);
     await frigateApp.page.getByLabel("Save Only").click();
     await expect
       .poll(() => capture.capturedUrl(), { timeout: 5_000 })
@@ -92,6 +94,7 @@ test.describe("Config Editor — Save @medium", () => {
     await expect(frigateApp.page.locator(".monaco-editor").first()).toBeVisible(
       { timeout: 15_000 },
     );
+    await makeMonacoEdit(frigateApp.page);
     await frigateApp.page.getByLabel("Save Only").click();
     await expect(frigateApp.page.getByText(/Invalid field/i)).toBeVisible({
       timeout: 5_000,
@@ -117,6 +120,7 @@ test.describe("Config Editor — Save and Restart @medium", () => {
       { timeout: 15_000 },
     );
 
+    await makeMonacoEdit(frigateApp.page);
     await frigateApp.page.getByLabel("Save & Restart").click();
     const dialog = frigateApp.page.getByRole("alertdialog");
     await expect(dialog).toBeVisible({ timeout: 5_000 });
@@ -140,6 +144,7 @@ test.describe("Config Editor — Save and Restart @medium", () => {
       { timeout: 15_000 },
     );
 
+    await makeMonacoEdit(frigateApp.page);
     await frigateApp.page.getByLabel("Save & Restart").click();
     const dialog = frigateApp.page.getByRole("alertdialog");
     await expect(dialog).toBeVisible({ timeout: 5_000 });

web/e2e/specs/settings/genai.spec.ts

@@ -0,0 +1,216 @@
+/**
+ * Generative AI provider settings tests -- MEDIUM tier.
+ *
+ * A model name belongs to its provider, so switching provider clears the model
+ * field. The roles widget strips a role only for a model or provider picked in
+ * the form, never when capability data arrives for the saved entry, which would
+ * dirty the section on load and silently drop the role on the next save.
+ */
+
+import { readFileSync } from "node:fs";
+import { resolve, dirname } from "node:path";
+import { fileURLToPath } from "node:url";
+import { test, expect } from "../../fixtures/frigate-test";
+import type { Page } from "@playwright/test";
+import { configFactory } from "../../fixtures/mock-data/config";
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const CONFIG_SCHEMA = JSON.parse(
+  readFileSync(
+    resolve(__dirname, "../../fixtures/mock-data/config-schema.json"),
+    "utf-8",
+  ),
+);
+
+const ENTRY = "audio";
+const SETTINGS_URL = "/settings?page=integrationGenerativeAi";
+const UNSAVED = "You have unsaved changes";
+const MODEL_PLACEHOLDER = "Select or enter a model…";
+
+type Entry = {
+  provider: string;
+  model: string;
+  base_url?: string;
+  roles: string[];
+};
+
+type ProviderInfo = {
+  models: string[];
+  supports_transcription: boolean;
+  model_capabilities?: Record<string, { supports_transcription?: boolean }>;
+};
+
+async function installRoutes(page: Page, entry: Entry, info: ProviderInfo) {
+  const config = configFactory({ genai: { [ENTRY]: entry } });
+
+  await page.route("**/api/config/schema.json", (route) =>
+    route.fulfill({ json: CONFIG_SCHEMA }),
+  );
+  await page.route("**/api/config", (route) => {
+    if (route.request().method() === "GET") {
+      return route.fulfill({ json: config });
+    }
+    return route.fulfill({ json: { success: true } });
+  });
+  await page.route("**/api/config/raw_paths", (route) =>
+    route.fulfill({ json: { genai: { [ENTRY]: entry } } }),
+  );
+  await page.route("**/api/genai/models", (route) =>
+    route.fulfill({
+      json: {
+        [ENTRY]: {
+          roles: entry.roles,
+          supports_toggleable_thinking: false,
+          supports_embeddings: true,
+          model_capabilities: {},
+          ...info,
+        },
+      },
+    }),
+  );
+}
+
+function roleSwitch(page: Page, role: string) {
+  return page.locator(`#root_${ENTRY}_roles-${role}`);
+}
+
+test.describe("genai provider settings @medium", () => {
+  test("a saved role the provider cannot confirm stays and is not dirty", async ({
+    frigateApp,
+  }) => {
+    // The server does not serve the saved model, so the backend reports every
+    // capability as false for the entry.
+    await installRoutes(
+      frigateApp.page,
+      {
+        provider: "llamacpp",
+        model: "stale-model",
+        base_url: "http://llama:8080",
+        roles: ["transcribe"],
+      },
+      { models: ["qwen3-asr"], supports_transcription: false },
+    );
+    await frigateApp.goto(SETTINGS_URL);
+
+    await expect(roleSwitch(frigateApp.page, "transcribe")).toBeVisible();
+    await expect(roleSwitch(frigateApp.page, "transcribe")).toBeChecked();
+
+    // Give any stripping effect time to fire, then confirm the section stayed
+    // clean.
+    await frigateApp.page.waitForTimeout(1000);
+    await expect(frigateApp.page.getByText(UNSAVED)).toBeHidden();
+    await expect(roleSwitch(frigateApp.page, "transcribe")).toBeChecked();
+  });
+
+  test("switching provider clears the model", async ({ frigateApp }) => {
+    await installRoutes(
+      frigateApp.page,
+      {
+        provider: "openai",
+        model: "gpt-4o",
+        roles: ["descriptions"],
+      },
+      { models: ["gpt-4o"], supports_transcription: true },
+    );
+    await frigateApp.goto(SETTINGS_URL);
+
+    const model = frigateApp.page.locator(`#root_${ENTRY}_model`);
+    await expect(model).toHaveText("gpt-4o");
+
+    await frigateApp.page.locator(`#root_${ENTRY}_provider`).click();
+    await frigateApp.page.getByRole("option", { name: "llamacpp" }).click();
+
+    await expect(model).toHaveText(MODEL_PLACEHOLDER);
+    await expect(frigateApp.page.getByText(UNSAVED)).toBeVisible();
+  });
+
+  test("switching back to the saved provider drops the other provider's model", async ({
+    frigateApp,
+  }) => {
+    await installRoutes(
+      frigateApp.page,
+      {
+        provider: "openai",
+        model: "gpt-4o",
+        roles: ["descriptions"],
+      },
+      { models: ["gpt-4o", "qwen3"], supports_transcription: true },
+    );
+    await frigateApp.goto(SETTINGS_URL);
+
+    const model = frigateApp.page.locator(`#root_${ENTRY}_model`);
+    const provider = frigateApp.page.locator(`#root_${ENTRY}_provider`);
+    await expect(model).toHaveText("gpt-4o");
+
+    await provider.click();
+    await frigateApp.page.getByRole("option", { name: "llamacpp" }).click();
+    await model.click();
+    await frigateApp.page.getByRole("option", { name: "qwen3" }).click();
+    await expect(model).toHaveText("qwen3");
+
+    // the model picked for llamacpp must not carry over to openai, and the
+    // saved one isn't filled back in since the endpoint may have changed
+    await provider.click();
+    await frigateApp.page
+      .getByRole("option", { name: "openai", exact: true })
+      .click();
+
+    await expect(model).toHaveText(MODEL_PLACEHOLDER);
+  });
+
+  test("undo after switching provider brings back the saved model", async ({
+    frigateApp,
+  }) => {
+    await installRoutes(
+      frigateApp.page,
+      {
+        provider: "openai",
+        model: "gpt-4o",
+        roles: ["descriptions"],
+      },
+      { models: ["gpt-4o"], supports_transcription: true },
+    );
+    await frigateApp.goto(SETTINGS_URL);
+
+    const model = frigateApp.page.locator(`#root_${ENTRY}_model`);
+    await frigateApp.page.locator(`#root_${ENTRY}_provider`).click();
+    await frigateApp.page.getByRole("option", { name: "llamacpp" }).click();
+    await expect(model).toHaveText(MODEL_PLACEHOLDER);
+
+    await frigateApp.page.getByRole("button", { name: "Undo" }).click();
+
+    await expect(model).toHaveText("gpt-4o");
+    await expect(frigateApp.page.getByText(UNSAVED)).toBeHidden();
+  });
+
+  test("picking a model that cannot transcribe strips the role", async ({
+    frigateApp,
+  }) => {
+    await installRoutes(
+      frigateApp.page,
+      {
+        provider: "llamacpp",
+        model: "qwen3-asr",
+        base_url: "http://llama:8080",
+        roles: ["transcribe"],
+      },
+      {
+        models: ["qwen3-asr", "text-only"],
+        supports_transcription: true,
+        model_capabilities: {
+          "qwen3-asr": { supports_transcription: true },
+          "text-only": { supports_transcription: false },
+        },
+      },
+    );
+    await frigateApp.goto(SETTINGS_URL);
+
+    await expect(roleSwitch(frigateApp.page, "transcribe")).toBeChecked();
+
+    await frigateApp.page.locator(`#root_${ENTRY}_model`).click();
+    await frigateApp.page.getByRole("option", { name: "text-only" }).click();
+
+    await expect(roleSwitch(frigateApp.page, "transcribe")).toBeHidden();
+    await expect(frigateApp.page.getByText(UNSAVED)).toBeVisible();
+  });
+});

web/e2e/specs/settings/runtime-overrides.spec.ts

@@ -0,0 +1,117 @@
+/**
+ * Runtime override tests -- MEDIUM tier.
+ *
+ * Live view, MQTT, and Home Assistant toggles change a camera's running config
+ * without touching yaml, and the change persists across restarts. The settings
+ * form edits yaml, so it keeps showing the saved value. Without a marker on the
+ * field and runtime-aware wording on dependent warnings, the two read as a
+ * contradiction: "audio detection is not enabled" next to a switch that is on.
+ */
+
+import { readFileSync } from "node:fs";
+import { resolve, dirname } from "node:path";
+import { fileURLToPath } from "node:url";
+import { test, expect } from "../../fixtures/frigate-test";
+import type { Page } from "@playwright/test";
+import { configFactory } from "../../fixtures/mock-data/config";
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const CONFIG_SCHEMA = JSON.parse(
+  readFileSync(
+    resolve(__dirname, "../../fixtures/mock-data/config-schema.json"),
+    "utf-8",
+  ),
+);
+
+const CAMERA = "front_door";
+const TRANSCRIPTION_URL = `/settings?page=cameraAudioTranscription&camera=${CAMERA}`;
+const AUDIO_URL = `/settings?page=cameraAudioEvents&camera=${CAMERA}`;
+
+const CONFIG_DISABLED = /Audio detection is not enabled for this camera/;
+const RUNTIME_DISABLED =
+  /Audio detection is enabled in your config, but it is currently turned off/;
+
+type AudioState = { enabled: boolean; enabled_in_config: boolean };
+
+async function installRoutes(page: Page, audio: AudioState) {
+  const config = configFactory({
+    cameras: {
+      [CAMERA]: {
+        audio,
+        // audio detection only runs on a stream carrying the audio role
+        ffmpeg: {
+          inputs: [
+            {
+              path: "rtsp://user:pass@host/front",
+              roles: ["record", "detect", "audio"],
+            },
+          ],
+        },
+        audio_transcription: { enabled: true, enabled_in_config: true },
+      },
+    },
+  });
+
+  await page.route("**/api/config/raw_paths", (route) =>
+    route.fulfill({
+      json: { cameras: { [CAMERA]: { ffmpeg: { inputs: [] } } } },
+    }),
+  );
+  await page.route("**/api/config/schema.json", (route) =>
+    route.fulfill({ json: CONFIG_SCHEMA }),
+  );
+  await page.route("**/api/config", (route) => {
+    if (route.request().method() === "GET") {
+      return route.fulfill({ json: config });
+    }
+    return route.fulfill({ json: { success: true } });
+  });
+}
+
+test.describe("runtime overrides @medium", () => {
+  test("a runtime-only toggle gets its own wording and a field marker", async ({
+    frigateApp,
+  }) => {
+    await installRoutes(frigateApp.page, {
+      enabled: false,
+      enabled_in_config: true,
+    });
+
+    await frigateApp.goto(TRANSCRIPTION_URL);
+    await expect(frigateApp.page.getByText(RUNTIME_DISABLED)).toBeVisible();
+    await expect(frigateApp.page.getByText(CONFIG_DISABLED)).toBeHidden();
+
+    // The audio section still shows the saved value, so the switch stays on and
+    // the marker carries the live state.
+    await frigateApp.goto(AUDIO_URL);
+    await expect(
+      frigateApp.page.getByRole("switch", { name: "Enable audio detection" }),
+    ).toHaveAttribute("data-state", "checked");
+    // the boolean layout renders a mobile and a desktop label block, so only
+    // one of the two badges is on screen
+    await expect(
+      frigateApp.page.getByText("Overridden (Live)").filter({ visible: true }),
+    ).toHaveCount(1);
+  });
+
+  test("a config-disabled section keeps the original wording", async ({
+    frigateApp,
+  }) => {
+    await installRoutes(frigateApp.page, {
+      enabled: false,
+      enabled_in_config: false,
+    });
+
+    await frigateApp.goto(TRANSCRIPTION_URL);
+    await expect(frigateApp.page.getByText(CONFIG_DISABLED)).toBeVisible();
+    await expect(frigateApp.page.getByText(RUNTIME_DISABLED)).toBeHidden();
+
+    await frigateApp.goto(AUDIO_URL);
+    await expect(
+      frigateApp.page.getByRole("switch", { name: "Enable audio detection" }),
+    ).toHaveAttribute("data-state", "unchecked");
+    await expect(
+      frigateApp.page.getByText("Overridden (Live)").filter({ visible: true }),
+    ).toHaveCount(0);
+  });
+});

web/e2e/specs/system.spec.ts

@@ -72,7 +72,15 @@ test.describe("System — tabs @medium", () => {
       { timeout: 15_000 },
     );
     await expect(frigateApp.page.getByText("0.15.0-test")).toBeVisible();
-    await expect(frigateApp.page.getByText(/Last refreshed/)).toBeVisible();
+
+    if (frigateApp.isMobile) {
+      // the "Last refreshed" label is dropped on mobile so the timestamp
+      // clears the centered logo
+      await expect(frigateApp.page.getByText(/Last refreshed/)).toHaveCount(0);
+      await expect(frigateApp.page.getByText(/Just now|ago/)).toBeVisible();
+    } else {
+      await expect(frigateApp.page.getByText(/Last refreshed/)).toBeVisible();
+    }
   });
 
   test("storage tab renders content after switching", async ({
@@ -234,4 +242,45 @@ test.describe("System — mobile @medium @mobile", () => {
       { timeout: 5_000 },
     );
   });
+
+  test("header controls leave the logo uncovered on a narrow phone", async ({
+    frigateApp,
+  }) => {
+    await frigateApp.goto("/system#general");
+    await expect(frigateApp.page.getByLabel("Select general")).toHaveAttribute(
+      "data-state",
+      "on",
+      { timeout: 15_000 },
+    );
+    await frigateApp.page.setViewportSize({ width: 320, height: 740 });
+
+    const logo = frigateApp.page.locator("svg.fill-current").first();
+    const tabs = frigateApp.page
+      .locator("[data-radix-scroll-area-viewport]")
+      .filter({ has: frigateApp.page.getByLabel("Select general") });
+    const refreshed = frigateApp.page.getByText(/Just now|ago/);
+
+    const logoBox = await logo.boundingBox();
+    const tabsBox = await tabs.boundingBox();
+    const refreshedBox = await refreshed.boundingBox();
+
+    expect(tabsBox!.x + tabsBox!.width).toBeLessThanOrEqual(logoBox!.x + 1);
+    expect(refreshedBox!.x).toBeGreaterThanOrEqual(logoBox!.x + logoBox!.width);
+
+    // the clipped tabs stay reachable by scrolling
+    const overflow = await tabs.evaluate((el) => ({
+      scroll: el.scrollWidth,
+      client: el.clientWidth,
+    }));
+    expect(overflow.scroll).toBeGreaterThan(overflow.client);
+    await tabs.evaluate((el) => {
+      el.scrollLeft = el.scrollWidth;
+    });
+    await frigateApp.page.getByLabel("Select cameras").click();
+    await expect(frigateApp.page.getByLabel("Select cameras")).toHaveAttribute(
+      "data-state",
+      "on",
+      { timeout: 5_000 },
+    );
+  });
 });

web/public/locales/ar/components/player.json

@@ -15,10 +15,6 @@
     "bandwidth": {
       "title": "العرض الترددي:",
       "short": "العرض الترددي"
-    },
-    "latency": {
-      "title": "التأخير:",
-      "value": "{{seconds}} ثانية"
     }
   },
   "streamOffline": {

web/public/locales/be/common.json

@@ -21,8 +21,8 @@
     "1hour": "1 гадзіна",
     "12hours": "12 гадзін",
     "24hours": "24 гадзіны",
-    "pm": "вечара",
-    "am": "раніцы",
+    "pm": "pm",
+    "am": "am",
     "yr": "{{time}} г",
     "year_one": "{{time}} год",
     "year_few": "{{time}} гады",
@@ -168,13 +168,16 @@
     "saveAll": "Захаваць усё",
     "savingAll": "Захаванне ўсяго…",
     "undoAll": "Адрабіць усё",
-    "retry": "Паўтарыць"
+    "retry": "Паўтарыць",
+    "expand": "Разгарнуць",
+    "collapse": "Згарнуць",
+    "reload": "Перазагрузіць"
   },
   "menu": {
     "system": "Сістэма",
-    "systemMetrics": "Сістэмныя метрыкі",
+    "systemMetrics": "Стан і метрыкі",
     "configuration": "Канфігурацыя",
-    "systemLogs": "Журналы сістэмы",
+    "systemLogs": "Журналы",
     "profiles": "Профілі",
     "settings": "Налады",
     "configurationEditor": "Рэдактар канфігурацыі",
@@ -225,7 +228,7 @@
       "withSystem": {
         "label": "Выкарыстоўваць мову з налад сістэмы"
       },
-      "be": "Беларуская"
+      "be": "Беларуская (беларуская)"
     },
     "appearance": "Выгляд",
     "darkMode": {
@@ -328,5 +331,36 @@
   "validation_errors": "Памылкі праверкі",
   "credentialField": {
     "savedPlaceholder": "Захавана - пакіньце пустым, каб не мяняць"
+  },
+  "error": {
+    "title": "Гэтая старонка перастала працаваць",
+    "desc": "На старонцы адбылася нечаканая памылка. Звычайна дапамагае перазагрузка. Калі гэта паўтараецца, скапіюйце падрабязнасці і дадайце іх да абмеркавання на GitHub.",
+    "staleTitle": "Патрэбна перазагрузка",
+    "staleDesc": "Частка інтэрфейсу не загрузілася - звычайна гэта значыць, што Frigate абнавіўся, пакуль гэтая ўкладка была адкрытая. Перазагрузіце, каб атрымаць новую версію.",
+    "partial": "Частка інтэрфейсу перастала працаваць.",
+    "copyDetails": "Скапіяваць падрабязнасці",
+    "copyFailed": "Не ўдалося скапіяваць падрабязнасці ў буфер абмену"
+  },
+  "commandMenu": {
+    "title": "Меню каманд",
+    "description": "Пошук старонак, камер, налад і дзеянняў",
+    "placeholder": "Пошук старонак, камер, налад і дзеянняў",
+    "empty": "Няма супадзенняў",
+    "hint": "{{key}} або / каб адкрыць, Enter каб выканаць",
+    "section": {
+      "recent": "Апошнія",
+      "pages": "Старонкі",
+      "cameras": "Камеры",
+      "cameraGroups": "Групы камер",
+      "settings": "Налады",
+      "actions": "Дзеянні"
+    },
+    "camera": {
+      "live": "Эфір",
+      "review": "Разгляд"
+    },
+    "action": {
+      "toggleTheme": "Пераключыць светлы і цёмны рэжым"
+    }
   }
 }

web/public/locales/be/components/dialog.json

@@ -137,6 +137,15 @@
       "queueingExport": "Экспарт ставіцца ў чаргу…",
       "previewExport": "Перадпрагляд экспарту",
       "useThisRange": "Ужыць гэты прамежак"
+    },
+    "stream": {
+      "label": "Якасць",
+      "auto": "Аўта",
+      "main": "Зыходная",
+      "sub": "Нізкая",
+      "autoDesc": "Выкарыстоўвае асноўную плынь, а дзе яна недаступная - пераходзіць на дадатковую плынь ніжэйшай якасці.",
+      "mainDesc": "Экспартуе толькі асноўную плынь зыходнай якасці. Прамежкі часу, дзе асноўная плынь недаступная, будуць адсутнічаць.",
+      "subDesc": "Экспартуе толькі дадатковую плынь ніжэйшай якасці. Прамежкі часу, дзе дадатковая плынь недаступная, будуць адсутнічаць."
     }
   },
   "streaming": {
@@ -190,7 +199,14 @@
       "export": "Экспарт",
       "markAsReviewed": "Пазначыць як разгледжанае",
       "markAsUnreviewed": "Пазначыць як неразгледжанае",
-      "deleteNow": "Выдаліць зараз"
+      "deleteNow": "Выдаліць зараз",
+      "generateDescription": "Стварыць апісанне"
+    },
+    "genaiDescription": {
+      "toast": {
+        "success": "Апісанне GenAI запытана для гэтага элемента разгляду.",
+        "error": "Не ўдалося запытаць апісанне: {{error}}"
+      }
     }
   },
   "imagePicker": {

web/public/locales/be/components/player.json

@@ -22,14 +22,6 @@
       "title": "Паласа прапускання:",
       "short": "Паласа"
     },
-    "latency": {
-      "title": "Затрымка:",
-      "value": "{{seconds}} сек",
-      "short": {
-        "title": "Затрымка",
-        "value": "{{seconds}} с"
-      }
-    },
     "totalFrames": "Усяго кадраў:",
     "droppedFrames": {
       "title": "Страчаныя кадры:",
@@ -46,7 +38,22 @@
       "submittedFrigatePlus": "Кадр паспяхова адпраўлены ў Frigate+"
     },
     "error": {
-      "submitFrigatePlusFailed": "Не ўдалося адправіць кадр у Frigate+"
+      "submitFrigatePlusFailed": "Не ўдалося адправіць кадр у Frigate+",
+      "playRecordingsFailed": "Не ўдалося прайграць запісы (памылка {{code}}): {{message}}"
     }
+  },
+  "quality": {
+    "auto": "Аўта",
+    "autoLow": "Прайграецца нізкая якасць (абмежаваная паласа прапускання)",
+    "autoLowCodec": "Прайграецца нізкая якасць (зыходная не падтрымліваецца гэтым браўзерам)",
+    "autoLowSaveData": "Прайграецца нізкая якасць (эканомія трафіку)",
+    "main": "Зыходная",
+    "sub": "Нізкая",
+    "notSupportedBrowser": "Не падтрымліваецца гэтым браўзерам",
+    "label": "Якасць",
+    "noAudio": "Без гуку",
+    "audioRate": "Гук {{rate}} кГц",
+    "audioCodecRate": "{{codec}} {{rate}} кГц",
+    "noRecordings": "Няма запісаў у гэтым прамежку часу"
   }
 }

web/public/locales/be/config/cameras.json

@@ -70,7 +70,7 @@
         },
         "objects": {
             "label": "Аб'екты-трыгеры",
-            "description": "Спіс тыпаў аб'ектаў (з labelmap), якія могуць спрацаваць у гэтай зоне. Можа быць радком або спісам радкоў. Калі пуста, улічваюцца ўсе аб'екты."
+            "description": "Спіс тыпаў аб'ектаў (з карты метак), якія могуць спрацаваць у гэтай зоне. Можа быць радком або спісам радкоў. Калі пуста, улічваюцца ўсе аб'екты."
         }
     },
     "label": "CameraConfig",
@@ -120,6 +120,10 @@
         "num_threads": {
             "label": "Патокі дэтэктавання",
             "description": "Колькасць патокаў для апрацоўкі дэтэктавання гуку."
+        },
+        "labelmap": {
+            "label": "Змяненне карты метак гуку",
+            "description": "Перавызначэнні або пераназначэнні, якія аб'ядноўваюцца са стандартнай картай метак гуку."
         }
     },
     "audio_transcription": {
@@ -147,6 +151,10 @@
         "order": {
             "label": "Пазіцыя",
             "description": "Лічбавае становішча, якое вызначае парадак камеры ў раскладцы Birdseye."
+        },
+        "modes": {
+            "label": "Тыпы актыўнасці",
+            "description": "Тыпы актыўнасці, пры якіх камеры паказваюцца ў Birdseye."
         }
     },
     "detect": {
@@ -207,6 +215,10 @@
         "annotation_offset": {
             "label": "Зрух анатацый",
             "description": "На колькі мілісекунд зрушыць прамавугольнікі аб'ектаў, каб яны супадалі з запісаным відэа: плыні «detect» і «record» рэдка ідэальна сінхронныя. Можа быць дадатным або адмоўным."
+        },
+        "scene": {
+            "label": "Сцэна дэтэктавання",
+            "description": "Асяроддзе, на якое глядзіць гэтая камера; паводле яго выбіраецца, якая з наладжаных мадэлей на ёй працуе. Камеры са значэннем «all» выкарыстоўваюць мадэль, наладжаную са сцэнай «all»."
         }
     },
     "face_recognition": {
@@ -250,6 +262,10 @@
             "record": {
                 "label": "Выходныя аргументы для record",
                 "description": "Прадвызначаныя выходныя аргументы для плыняў з роляй «record»."
+            },
+            "record_sub": {
+                "label": "Выходныя аргументы запісу дадатковай плыні",
+                "description": "Выходныя аргументы для плыняў з роляй «record_sub». Калі не зададзена, выкарыстоўваюцца выходныя аргументы запісу."
             }
         },
         "retry_interval": {
@@ -579,6 +595,54 @@
         "enabled_in_config": {
             "label": "Зыходны стан запісу",
             "description": "Паказвае, ці быў запіс уключаны ў зыходнай статычнай канфігурацыі."
+        },
+        "sub": {
+            "label": "Запіс дадатковай плыні",
+            "description": "Налады запісу другой плыні ніжэйшай якасці.",
+            "enabled": {
+                "label": "Уключыць запіс дадатковай плыні",
+                "description": "Уключыць запіс другой плыні ніжэйшай якасці для прайгравання з адаптыўнай якасцю і падоўжанага захоўвання."
+            },
+            "continuous": {
+                "label": "Бесперапыннае захоўванне дадатковай плыні",
+                "description": "Колькасць дзён захоўвання запісаў дадатковай плыні незалежна ад аб'ектаў пад адсочваннем або руху.",
+                "days": {
+                    "label": "Дні захоўвання",
+                    "description": "Дзён захоўвання запісаў."
+                }
+            },
+            "motion": {
+                "label": "Захоўванне дадатковай плыні па руху",
+                "description": "Колькасць дзён захоўвання запісаў дадатковай плыні, выкліканых рухам.",
+                "days": {
+                    "label": "Дні захоўвання",
+                    "description": "Дзён захоўвання запісаў."
+                }
+            },
+            "alerts": {
+                "label": "Захоўванне абвестак дадатковай плыні",
+                "description": "Налады захоўвання запісаў абвестак дадатковай плыні.",
+                "days": {
+                    "label": "Дні захоўвання",
+                    "description": "Колькасць дзён захоўвання запісаў падзей дэтэктавання."
+                },
+                "mode": {
+                    "label": "Рэжым захоўвання",
+                    "description": "Рэжым захоўвання: all (усе сегменты), motion (сегменты з рухам) або active_objects (сегменты з актыўнымі аб'ектамі)."
+                }
+            },
+            "detections": {
+                "label": "Захоўванне дэтэктаванняў дадатковай плыні",
+                "description": "Налады захоўвання запісаў дэтэктаванняў дадатковай плыні.",
+                "days": {
+                    "label": "Дні захоўвання",
+                    "description": "Колькасць дзён захоўвання запісаў падзей дэтэктавання."
+                },
+                "mode": {
+                    "label": "Рэжым захоўвання",
+                    "description": "Рэжым захоўвання: all (усе сегменты), motion (сегменты з рухам) або active_objects (сегменты з актыўнымі аб'ектамі)."
+                }
+            }
         }
     },
     "review": {
@@ -670,6 +734,10 @@
             "activity_context_prompt": {
                 "label": "Прампт кантэксту актыўнасці",
                 "description": "Уласны прампт, які апісвае, якая актыўнасць падазроная, а якая не, каб даць кантэкст для зводак GenAI."
+            },
+            "response_style": {
+                "label": "Стыль адказу",
+                "description": "Гатовы стыль пісьма для створаных апісанняў разгляду. Гатовыя стылі наладжваюць тон і падрабязнасць загалоўка, зводкі і апісання сцэны, якія бачыць карыстальнік; «default» пакідае ўбудаваны прампт без змен."
             }
         }
     },

web/public/locales/be/config/global.json

@@ -33,7 +33,11 @@
             "label": "Патокі дэтэктавання",
             "description": "Колькасць патокаў для апрацоўкі дэтэктавання гуку."
         },
-        "description": "Налады дэтэктавання падзей па гуку для ўсіх камер. Можна перавызначыць для кожнай камеры."
+        "description": "Налады дэтэктавання падзей па гуку для ўсіх камер. Можна перавызначыць для кожнай камеры.",
+        "labelmap": {
+            "label": "Змяненне карты метак гуку",
+            "description": "Перавызначэнні або пераназначэнні, якія аб'ядноўваюцца са стандартнай картай метак гуку."
+        }
     },
     "audio_transcription": {
         "label": "Транскрыпцыя гуку",
@@ -105,6 +109,10 @@
         "idle_heartbeat_fps": {
             "label": "FPS сігналу пры прастоі",
             "description": "Колькі разоў на секунду паўтараць апошні кадр Birdseye пры прастоі. 0 адключае."
+        },
+        "modes": {
+            "label": "Тыпы актыўнасці",
+            "description": "Тыпы актыўнасці, пры якіх камеры паказваюцца ў Birdseye."
         }
     },
     "detect": {
@@ -165,6 +173,10 @@
         "annotation_offset": {
             "label": "Зрух анатацый",
             "description": "На колькі мілісекунд зрушыць прамавугольнікі аб'ектаў, каб яны супадалі з запісаным відэа: плыні «detect» і «record» рэдка ідэальна сінхронныя. Можа быць дадатным або адмоўным."
+        },
+        "scene": {
+            "label": "Сцэна дэтэктавання",
+            "description": "Асяроддзе, на якое глядзіць гэтая камера; паводле яго выбіраецца, якая з наладжаных мадэлей на ёй працуе. Камеры са значэннем «all» выкарыстоўваюць мадэль, наладжаную са сцэнай «all»."
         }
     },
     "face_recognition": {
@@ -238,6 +250,10 @@
             "record": {
                 "label": "Выходныя аргументы для record",
                 "description": "Прадвызначаныя выходныя аргументы для плыняў з роляй «record»."
+            },
+            "record_sub": {
+                "label": "Выходныя аргументы запісу дадатковай плыні",
+                "description": "Выходныя аргументы для плыняў з роляй «record_sub». Калі не зададзена, выкарыстоўваюцца выходныя аргументы запісу."
             }
         },
         "retry_interval": {
@@ -650,7 +666,55 @@
             "label": "Зыходны стан запісу",
             "description": "Паказвае, ці быў запіс уключаны ў зыходнай статычнай канфігурацыі."
         },
-        "description": "Налады запісу і захоўвання для камер, калі яны не перавызначаны асобна."
+        "description": "Налады запісу і захоўвання для камер, калі яны не перавызначаны асобна.",
+        "sub": {
+            "label": "Запіс дадатковай плыні",
+            "description": "Налады запісу другой плыні ніжэйшай якасці.",
+            "enabled": {
+                "label": "Уключыць запіс дадатковай плыні",
+                "description": "Уключыць запіс другой плыні ніжэйшай якасці для прайгравання з адаптыўнай якасцю і падоўжанага захоўвання."
+            },
+            "continuous": {
+                "label": "Бесперапыннае захоўванне дадатковай плыні",
+                "description": "Колькасць дзён захоўвання запісаў дадатковай плыні незалежна ад аб'ектаў пад адсочваннем або руху.",
+                "days": {
+                    "label": "Дні захоўвання",
+                    "description": "Дзён захоўвання запісаў."
+                }
+            },
+            "motion": {
+                "label": "Захоўванне дадатковай плыні па руху",
+                "description": "Колькасць дзён захоўвання запісаў дадатковай плыні, выкліканых рухам.",
+                "days": {
+                    "label": "Дні захоўвання",
+                    "description": "Дзён захоўвання запісаў."
+                }
+            },
+            "alerts": {
+                "label": "Захоўванне абвестак дадатковай плыні",
+                "description": "Налады захоўвання запісаў абвестак дадатковай плыні.",
+                "days": {
+                    "label": "Дні захоўвання",
+                    "description": "Колькасць дзён захоўвання запісаў падзей дэтэктавання."
+                },
+                "mode": {
+                    "label": "Рэжым захоўвання",
+                    "description": "Рэжым захоўвання: all (усе сегменты), motion (сегменты з рухам) або active_objects (сегменты з актыўнымі аб'ектамі)."
+                }
+            },
+            "detections": {
+                "label": "Захоўванне дэтэктаванняў дадатковай плыні",
+                "description": "Налады захоўвання запісаў дэтэктаванняў дадатковай плыні.",
+                "days": {
+                    "label": "Дні захоўвання",
+                    "description": "Колькасць дзён захоўвання запісаў падзей дэтэктавання."
+                },
+                "mode": {
+                    "label": "Рэжым захоўвання",
+                    "description": "Рэжым захоўвання: all (усе сегменты), motion (сегменты з рухам) або active_objects (сегменты з актыўнымі аб'ектамі)."
+                }
+            }
+        }
     },
     "review": {
         "label": "Разгляд",
@@ -740,6 +804,10 @@
             "activity_context_prompt": {
                 "label": "Прампт кантэксту актыўнасці",
                 "description": "Уласны прампт, які апісвае, якая актыўнасць падазроная, а якая не, каб даць кантэкст для зводак GenAI."
+            },
+            "response_style": {
+                "label": "Стыль адказу",
+                "description": "Гатовы стыль пісьма для створаных апісанняў разгляду. Гатовыя стылі наладжваюць тон і падрабязнасць загалоўка, зводкі і апісання сцэны, якія бачыць карыстальнік; «default» пакідае ўбудаваны прампт без змен."
             }
         },
         "description": "Налады, якія кіруюць абвесткамі, дэтэктаваннямі і зводкамі разгляду GenAI для інтэрфейсу і сховішча."
@@ -1173,7 +1241,7 @@
         },
         "default_role": {
             "label": "Прадвызначаная роля",
-            "description": "Роля, якая прызначаецца карыстальнікам, аўтарызаваным праз проксі, калі супастаўленне роляў не спрацавала."
+            "description": "Роля, якая прызначаецца карыстальнікам, аўтарызаваным праз проксі, калі супастаўленне роляў не спрацавала. Задайце «none», каб адмовіць у доступе карыстальнікам без супастаўлення."
         },
         "separator": {
             "label": "Сімвал-раздзяляльнік",
@@ -1399,5 +1467,57 @@
             "label": "Паказваць у разглядзе",
             "description": "Ці бачная гэта камера ў разглядзе (старонка разгляду і яе фільтр камер, разгляд руху і выгляд гісторыі)."
         }
+    },
+    "models": {
+        "label": "Мадэлі дэтэктавання",
+        "description": "Мадэлі дэтэктавання аб'ектаў і абсталяванне, на якім працуе кожная з іх. Камера выбірае мадэль, супастаўляючы свой detect.scene са сцэнай мадэлі.",
+        "scene": {
+            "label": "Сцэна мадэлі",
+            "description": "Асяроддзе камеры, для якога прызначана гэтая мадэль. Камера выбірае мадэль, задаючы detect.scene з адпаведным значэннем, а «all» выкарыстоўвае любая камера, якая нічога не задала."
+        },
+        "devices": {
+            "label": "Абсталяванне дэтэктавання",
+            "description": "Абсталяванне, на якім працуе гэтая мадэль, у выглядзе «<detector>» або «<detector>:<device>» (напрыклад «edgetpu:pci:0» ці «openvino:GPU»). Калі адна прылада пазначана некалькі разоў, на ёй запускаюцца дадатковыя працэсы вывядзення."
+        },
+        "path": {
+            "label": "Шлях да мадэлі свайго дэтэктара аб'ектаў",
+            "description": "Шлях да файла свае мадэлі дэтэктавання (або plus://<model_id> для мадэлей Frigate+)."
+        },
+        "labelmap_path": {
+            "label": "Карта метак для свайго дэтэктара аб'ектаў",
+            "description": "Шлях да файла карты метак, які супастаўляе лікавыя класы з тэкставымі меткамі для дэтэктара."
+        },
+        "width": {
+            "label": "Шырыня ўваходу мадэлі дэтэктавання аб'ектаў",
+            "description": "Шырыня ўваходнага тэнзара мадэлі ў пікселях."
+        },
+        "height": {
+            "label": "Вышыня ўваходу мадэлі дэтэктавання аб'ектаў",
+            "description": "Вышыня ўваходнага тэнзара мадэлі ў пікселях."
+        },
+        "labelmap": {
+            "label": "Змяненне карты метак",
+            "description": "Перавызначэнні або пераназначэнні, якія аб'ядноўваюцца са стандартнай картай метак."
+        },
+        "attributes_map": {
+            "label": "Карта метак аб'ектаў да іх метак атрыбутаў",
+            "description": "Супастаўленне метак аб'ектаў з меткамі атрыбутаў, якія дадаюць метаданыя (напрыклад «car» -> ['license_plate'])."
+        },
+        "input_tensor": {
+            "label": "Форма ўваходнага тэнзара мадэлі",
+            "description": "Фармат тэнзара, якога чакае мадэль: «nhwc» або «nchw»."
+        },
+        "input_pixel_format": {
+            "label": "Фармат колеру пікселяў на ўваходзе мадэлі",
+            "description": "Каляровая прастора пікселяў, якой чакае мадэль: «rgb», «bgr» або «yuv»."
+        },
+        "input_dtype": {
+            "label": "Тып даных на ўваходзе мадэлі",
+            "description": "Тып даных уваходнага тэнзара мадэлі (напрыклад «float32»)."
+        },
+        "model_type": {
+            "label": "Тып мадэлі дэтэктавання аб'ектаў",
+            "description": "Тып архітэктуры мадэлі дэтэктара (ssd, yolox, yolonas, yolo-generic, rfdetr, dfine), які некаторыя дэтэктары выкарыстоўваюць для аптымізацыі."
+        }
     }
 }

web/public/locales/be/config/validation.json

@@ -31,5 +31,8 @@
     },
     "detect": {
         "dimensionMustBeEven": "Мусіць быць цотным лікам."
+    },
+    "models": {
+        "defaultRequired": "Адна мадэль мусіць выкарыстоўваць сцэну «All cameras». Без гэтага камера, якая не выбрала сцэну, не мае мадэлі, да якой можна вярнуцца."
     }
 }

web/public/locales/be/views/chat.json

@@ -55,6 +55,12 @@
         "auto_scroll": {
             "title": "Аўтапракрутка",
             "desc": "Сачыць за новымі паведамленнямі па меры паступлення."
+        },
+        "always_allow": {
+            "title": "Заўсёды дазволеныя дзеянні",
+            "desc": "Дзеянні, якія памочнік можа выконваць без папярэдняга запыту.",
+            "none": "Няма",
+            "reset": "Скінуць"
         }
     },
     "stats": {
@@ -68,5 +74,15 @@
     },
     "thinking": {
         "toggle": "Пераключыць разважанне"
+    },
+    "approval": {
+        "title": "Ухваліць {{tool}}?",
+        "desc": "Гэтае дзеянне нешта змяняе ў Frigate. Праглядзіце падрабязнасці перад ухваленнем.",
+        "approve": "Ухваліць",
+        "always_allow": "Заўсёды дазваляць",
+        "reject": "Адхіліць",
+        "approved": "Ухвалена",
+        "rejected": "Адхілена",
+        "placeholder": "Ухваліце або адхіліце дзеянне ў чаканні, каб працягнуць"
     }
 }

web/public/locales/be/views/events.json

@@ -91,5 +91,6 @@
     "crop": "Абрэзаць па фільтры",
     "cropAria": "Пераключыць абрэзку перадпрагляду па выбраных вобласцях",
     "cropDesc": "Набліжаць перадпрагляды да выбраных вобласцей фільтра замест паказу ўсяго кадра."
-  }
+  },
+  "subOnlyQuality": "У гэтым прамежку часу даступныя толькі запісы нізкай якасці"
 }

web/public/locales/be/views/exports.json

@@ -63,7 +63,8 @@
     "copying": "Капіяванне",
     "encoding": "Кадаванне",
     "encodingRetry": "Кадаванне (паўтор)",
-    "finalizing": "Завяршэнне"
+    "finalizing": "Завяршэнне",
+    "merging": "Аб'яднанне"
   },
   "caseView": {
     "noDescription": "Без апісання",

web/public/locales/be/views/faceLibrary.json

@@ -80,9 +80,9 @@
       "deletedFace_one": "Выдалены {{count}} твар.",
       "deletedFace_few": "Выдалена {{count}} твары.",
       "deletedFace_many": "Выдалена {{count}} твараў.",
-      "deletedName_one": "Выдалены {{count}} твар.",
-      "deletedName_few": "Выдалена {{count}} твары.",
-      "deletedName_many": "Выдалена {{count}} твараў.",
+      "deletedName_one": "{{count}} твар паспяхова выдалены.",
+      "deletedName_few": "{{count}} твары паспяхова выдалены.",
+      "deletedName_many": "{{count}} твараў паспяхова выдалена.",
       "renamedFace": "Твар перайменаваны на {{name}}",
       "trainedFace": "Твар навучаны.",
       "reclassifiedFace": "Твар перакласіфікаваны.",

web/public/locales/be/views/live.json

@@ -6,7 +6,12 @@
   "lowBandwidthMode": "Рэжым нізкай паласы прапускання",
   "twoWayTalk": {
     "enable": "Уключыць двухбаковую размову",
-    "disable": "Адключыць двухбаковую размову"
+    "disable": "Адключыць двухбаковую размову",
+    "requiresWebRTC": "Двухбаковая размова патрабуе WebRTC, які недаступны",
+    "error": {
+      "microphone": "Двухбаковая размова не атрымала доступу да вашага мікрафона",
+      "refused": "Не ўдалося пачаць двухбаковую размову. Падрабязнасці глядзіце ў кансолі браўзера."
+    }
   },
   "cameraAudio": {
     "enable": "Уключыць гук камеры",
@@ -135,7 +140,8 @@
       "unavailable": "Для гэтай плыні гук недаступны"
     },
     "debug": {
-      "picker": "Выбар плыні недаступны ў рэжыме адладкі. Адладачны выгляд заўсёды выкарыстоўвае плынь з роляй «detect»."
+      "picker": "Выбар плыні недаступны ў рэжыме адладкі. Адладачны выгляд заўсёды выкарыстоўвае плынь з роляй «detect».",
+      "technology": "Выбар тэхналогіі трансляцыі недаступны ў рэжыме адладкі."
     },
     "twoWayTalk": {
       "tips": "Ваша прылада мусіць падтрымліваць гэтую функцыю, а WebRTC мусіць быць наладжаны для двухбаковай размовы.",
@@ -144,11 +150,36 @@
     },
     "lowBandwidth": {
       "tips": "Жывы прагляд працуе ў рэжыме нізкай паласы прапускання праз буферызацыю або памылкі плыні.",
-      "resetStream": "Скінуць плынь"
+      "resetStream": "Скінуць плынь",
+      "force": {
+        "label": "Прымусова ўключыць рэжым нізкай паласы прапускання",
+        "desc": "Заўсёды прайграваць убудаваную ў Frigate плынь нізкай паласы прапускання замест выбранай. Працуе пры любым злучэнні, але мае ніжэйшую якасць і без гуку."
+      }
     },
     "playInBackground": {
       "label": "Прайграваць у фоне",
       "tips": "Уключыце гэты параметр, каб плынь працягваўся, калі плэер схаваны."
+    },
+    "mode": "Тэхналогія трансляцыі",
+    "technology": {
+      "description": "Выберыце пажаданую тэхналогію трансляцыі. Пры памылках прайгравання ці сеткі Frigate усё роўна можа перайсці ў рэжым нізкай паласы прапускання.",
+      "name": {
+        "mse": "MSE",
+        "webrtc": "WebRTC",
+        "jsmpeg": "JSMpeg"
+      },
+      "tips": {
+        "mse": "Рэкамендаваны прадвызначаны варыянт, шырокая сумяшчальнасць і плыўнае прайграванне",
+        "webrtc": "Патрабуе дадатковай наладкі і падтрымліваецца не на кожнай прыладзе"
+      },
+      "unavailable": {
+        "browser": "Ваш браўзер не падтрымлівае WebRTC.",
+        "not-configured": "WebRTC не наладжаны. Задайце ў go2rtc webrtc candidates або ice_servers.",
+        "unreachable": "WebRTC не змог злучыцца. Праверце, ці даступны порт 8555 і ці правільна зададзены серверы STUN/TURN.",
+        "video-codec": "Відэакодэк гэтай плыні не падтрымліваецца WebRTC у гэтым браўзеры.",
+        "audio-codec": "Аўдыякодэк гэтай плыні не падтрымліваецца WebRTC. Перакадуйце ў opus або G.711 праз go2rtc.",
+        "checking": "Праверка даступнасці WebRTC…"
+      }
     }
   },
   "cameraSettings": {

web/public/locales/be/views/settings.json

@@ -29,7 +29,7 @@
     "systemTelemetry": "Тэлеметрыя",
     "systemBirdseye": "Birdseye",
     "systemFfmpeg": "FFMPEG",
-    "systemDetectorsAndModel": "Дэтэктары і мадэль",
+    "systemDetectorsAndModel": "Мадэлі дэтэктавання",
     "systemMqtt": "MQTT",
     "systemGo2rtcStreams": "Плыні go2rtc",
     "integrationSemanticSearch": "Семантычны пошук",
@@ -204,11 +204,50 @@
     "toast": {
       "success": {
         "clearStoredLayout": "Захаваная раскладка для {{cameraName}} ачышчана",
-        "clearStreamingSettings": "Налады трансляцыі для ўсіх груп камер ачышчаны."
+        "clearStreamingSettings": "Налады трансляцыі для ўсіх груп камер ачышчаны.",
+        "exportUiSettings": "Налады экспартаваны",
+        "importUiSettings": "Налады імпартаваны"
       },
       "error": {
         "clearStoredLayoutFailed": "Не ўдалося ачысціць захаваную раскладку: {{errorMessage}}",
-        "clearStreamingSettingsFailed": "Не ўдалося ачысціць налады трансляцыі: {{errorMessage}}"
+        "clearStreamingSettingsFailed": "Не ўдалося ачысціць налады трансляцыі: {{errorMessage}}",
+        "exportUiSettingsFailed": "Не ўдалося экспартаваць налады",
+        "importNothingToApply": "Не ўдалося імпартаваць налады: файл не змяшчае налад",
+        "importInvalidJson": "Не ўдалося імпартаваць налады: файл не з'яўляецца сапраўдным JSON",
+        "importWrongType": "Не ўдалося імпартаваць налады: файл не з'яўляецца экспартам налад Frigate",
+        "importUnsupportedVersion": "Не ўдалося імпартаваць налады: файлу патрэбна навейшая версія Frigate",
+        "importInvalidSchema": "Не ўдалося імпартаваць налады: файл пашкоджаны",
+        "importUiSettingsFailed": "Не ўдалося імпартаваць налады"
+      }
+    },
+    "backupRestore": {
+      "title": "Рэзервовая копія і аднаўленне",
+      "transfer": {
+        "label": "Файл налад прылады",
+        "desc": "Раскладкі груп камер, налады трансляцыі і параметры інтэрфейсу захоўваюцца ў вашым браўзеры. Экспартуйце іх у файл, каб зрабіць рэзервовую копію або перанесці на іншую прыладу.",
+        "export": "Экспартаваць налады",
+        "import": "Імпартаваць налады"
+      },
+      "importDialog": {
+        "title": "Імпартаваць налады",
+        "desc": "Выберыце, што прымяніць з гэтага файла. Пасля імпарту Frigate перазагрузіцца.",
+        "exportedFrom": "Экспартавана {{date}} з канфігурацыі Frigate версіі {{version}}",
+        "layouts_one": "Раскладкі груп камер ({{count}} група)",
+        "layouts_few": "Раскладкі груп камер ({{count}} групы)",
+        "layouts_many": "Раскладкі груп камер ({{count}} груп)",
+        "streaming_one": "Налады трансляцыі ({{count}} камера)",
+        "streaming_few": "Налады трансляцыі ({{count}} камеры)",
+        "streaming_many": "Налады трансляцыі ({{count}} камер)",
+        "preferences_one": "Параметры інтэрфейсу ({{count}} налада)",
+        "preferences_few": "Параметры інтэрфейсу ({{count}} налады)",
+        "preferences_many": "Параметры інтэрфейсу ({{count}} налад)",
+        "unknownGroups_one": "Групы камер {{groups}} няма на гэтым серверы. Яе налады захаваюцца, але не будуць выкарыстоўвацца.",
+        "unknownGroups_few": "Груп камер {{groups}} няма на гэтым серверы. Іх налады захаваюцца, але не будуць выкарыстоўвацца.",
+        "unknownGroups_many": "Груп камер {{groups}} няма на гэтым серверы. Іх налады захаваюцца, але не будуць выкарыстоўвацца.",
+        "unknownCameras_one": "Камеры {{cameras}} няма на гэтым серверы. Яе налады трансляцыі будуць ігнаравацца.",
+        "unknownCameras_few": "Камер {{cameras}} няма на гэтым серверы. Іх налады трансляцыі будуць ігнаравацца.",
+        "unknownCameras_many": "Камер {{cameras}} няма на гэтым серверы. Іх налады трансляцыі будуць ігнаравацца.",
+        "confirm": "Імпартаваць"
       }
     }
   },
@@ -370,7 +409,7 @@
       }
     },
     "step3": {
-      "description": "Наладзьце ролі плыняў і дадайце дадатковыя плыні для вашай камеры.",
+      "description": "Наладзьце магчымасці і ролі плыняў, дадайце дадатковыя плыні для вашай камеры.",
       "streamsTitle": "Плыні камеры",
       "addStream": "Дадаць плынь",
       "addAnotherStream": "Дадаць яшчэ плынь",
@@ -404,11 +443,26 @@
         "title": "Ролі плыняў",
         "detect": "Асноўная плынь для дэтэктавання аб'ектаў.",
         "record": "Захоўвае сегменты відэаплыні паводле налад канфігурацыі.",
-        "audio": "Плынь для дэтэктавання па гуку."
+        "audio": "Плынь для дэтэктавання па гуку.",
+        "record_sub": "Запісы ніжэйшай якасці для адаптыўнага прайгравання і падоўжанага захоўвання (звычайна дадатковая плынь камеры)."
       },
       "featuresPopover": {
         "title": "Магчымасці плыні",
         "description": "Скарыстайцеся рэтрансляцыяй go2rtc, каб паменшыць колькасць падлучэнняў да камеры."
+      },
+      "appleCompatibility": {
+        "title": "Палепшыць прайграванне на прыладах Apple",
+        "description": "Уключыце, калі глядзіце запісы ў Safari або на iPhone, iPad ці Mac."
+      },
+      "ptz": {
+        "title": "Уключыць кіраванне PTZ",
+        "detectedNote": "Праз ONVIF выяўлена падтрымка PTZ. Калі гэта камера PTZ, з гэтым параметрам вы зможаце кіраваць паваротам, нахілам і маштабаваннем камеры з інтэрфейсу.",
+        "connectionDetails": "Падрабязнасці злучэння ONVIF",
+        "host": "Хост ONVIF",
+        "port": "Порт ONVIF",
+        "username": "Імя карыстальніка ONVIF",
+        "password": "Пароль ONVIF",
+        "hostRequiredWarning": "Калі ўключана кіраванне PTZ, патрэбны хост і порт ONVIF."
       }
     },
     "step4": {
@@ -1435,7 +1489,9 @@
         "orphansDeleted": "Выдалена асірацелых",
         "aborted": "Перарвана. Выдаленне перавысіла б парог бяспекі.",
         "error": "Памылка",
-        "totals": "Усяго"
+        "totals": "Усяго",
+        "spaceToReclaim": "Месца для вызвалення",
+        "spaceReclaimed": "Вызвалена месца"
       },
       "event_snapshots": "Здымкі аб'ектаў пад адсочваннем",
       "event_thumbnails": "Паменшаныя выявы аб'ектаў пад адсочваннем",
@@ -1488,7 +1544,12 @@
     },
     "knownPlates": {
       "namePlaceholder": "напрыклад Аўто жонкі",
-      "platePlaceholder": "Нумар знака або рэгулярны выраз"
+      "platePlaceholder": "Нумар знака або рэгулярны выраз",
+      "assignedTo": "Прызначана: {{name}}",
+      "detected": "Выяўленыя нумарныя знакі",
+      "noneDetected": "Нумарных знакаў пакуль не выяўлена",
+      "search": "Знайдзіце або ўвядзіце нумарны знак",
+      "useCustom": "Выкарыстаць «{{value}}»"
     },
     "liveStreams": {
       "streamNameLabel": "Назва плыні",
@@ -1544,7 +1605,8 @@
         "preset-record-mjpeg": "Запіс - камеры MJPEG",
         "preset-record-jpeg": "Запіс - камеры JPEG",
         "preset-record-ubiquiti": "Запіс - камеры Ubiquiti"
-      }
+      },
+      "sameAsRecord": "Тое самае, што выходныя аргументы запісу"
     },
     "cameraInputs": {
       "itemTitle": "Плынь {{index}}",
@@ -1629,8 +1691,11 @@
       "options": {
         "detect": "Дэтэктаваць",
         "record": "Запіс",
-        "audio": "Аўдыя"
-      }
+        "audio": "Аўдыя",
+        "record_sub": "Запіс (дадатковая плынь)"
+      },
+      "roleInUse": "Ужо прызначана іншай плыні",
+      "recordSubConflict": "Плынь не можа мець адначасова ролі record і record_sub"
     },
     "genaiRoles": {
       "options": {
@@ -1683,7 +1748,18 @@
     },
     "defaultRole": {
       "admin": "Адміністратар",
-      "viewer": "Назіральнік"
+      "viewer": "Назіральнік",
+      "none": "Няма (адмовіць у доступе)"
+    },
+    "birdseyeModes": {
+      "summary": "Выбрана: {{count}}",
+      "options": {
+        "continuous": "Бесперапынна",
+        "motion": "Рух",
+        "all_objects": "Усе аб'екты",
+        "alerts": "Абвесткі",
+        "detections": "Дэтэктаванні"
+      }
     }
   },
   "globalConfig": {
@@ -1905,7 +1981,8 @@
       "maxFramesSet": "Задаванне максімуму кадраў перавызначае прадвызначаныя паводзіны і адключае адсочванне нерухомых аб'ектаў. Гэта патрэбна вельмі рэдка, карыстайцеся асцярожна.",
       "squareResolution": "Квадратнае разрозненне дэтэктавання - незвычайны выбар. Шырыня і вышыня дэтэктавання мусяць адпавядаць фарматным суадносінам камеры (напрыклад 16:9), а не памерам мадэлі дэтэктавання аб'ектаў. Неадпаведнасць можа расцягнуць відарыс і знізіць дакладнасць дэтэктавання.",
       "resolutionHigh": "Гэта разрозненне дэтэктавання вышэйшае за рэкамендаванае і можа павялічыць спажыванне рэсурсаў без павышэння дакладнасці. Для большасці камер рэкамендуецца разрозненне не вышэй за 1080p.",
-      "globalResolutionMultipleCameras": "Зададзена глабальнае разрозненне дэтэктавання, хоць наладжана некалькі камер. Калі камеры не маюць аднолькавых разрознення і фарматных суадносін, шырыню і вышыню дэтэктавання варта задаваць для кожнай камеры асобна."
+      "globalResolutionMultipleCameras": "Зададзена глабальнае разрозненне дэтэктавання, хоць наладжана некалькі камер. Калі камеры не маюць аднолькавых разрознення і фарматных суадносін, шырыню і вышыню дэтэктавання варта задаваць для кожнай камеры асобна.",
+      "sceneWithoutModel": "Для гэтай сцэны не наладжана мадэль дэтэктавання, таму камера вяртаецца да мадэлі са сцэнай «All cameras». Дадайце мадэль для гэтай сцэны, каб камера мела сваю."
     },
     "model": {
       "optimizedFor320": "Frigate аптымізаваны пад мадэль 320x320, і для большасці канфігурацый гэта найлепшы выбар. Мадэль 640x640 працуе павольней і дапамагае толькі ў асобных выпадках.",
@@ -1929,7 +2006,8 @@
       "modelSizeLarge": "Мадэль «large» аптымізавана пад шматрадковыя нумарныя знакі. Мадэль «small» працуе хутчэй, і карыстацца варта ёй, калі ў вашым рэгіёне не ўжываюцца шматрадковыя знакі."
     },
     "record": {
-      "noRecordRole": "Ніводнай плыні не прызначана роля «record». Запіс працаваць не будзе."
+      "noRecordRole": "Ніводнай плыні не прызначана роля «record». Запіс працаваць не будзе.",
+      "noRecordSubRole": "Ні ў адной плыні не зададзена роля record_sub. Запіс дадатковай плыні не будзе працаваць."
     },
     "snapshots": {
       "detectDisabled": "Дэтэктаванне аб'ектаў адключана. Здымкі ствараюцца з аб'ектаў пад адсочваннем і стварацца не будуць."
@@ -1939,6 +2017,50 @@
     },
     "onvif": {
       "autotrackingNoZones": "Аўтаадсочванне патрабуе хаця б адной зоны. Вызначце зону для гэтай камеры ў раздзеле «Маскі і зоны», а потым пазначце яе як патрэбную ніжэй."
+    },
+    "birdseye": {
+      "objectTrackingDetectDisabled": "Birdseye паказвае аб'екты пад адсочваннем, але дэтэктаванне аб'ектаў для гэтай камеры адключана. Камера не з'явіцца ў Birdseye."
+    }
+  },
+  "detectionModels": {
+    "title": "Мадэлі дэтэктавання",
+    "description": "Наладзьце мадэлі дэтэктавання аб'ектаў і абсталяванне, на якім працуе кожная з іх. Камера выбірае мадэль паводле сваёй сцэны ў наладах дэтэктавання.",
+    "addModel": "Дадаць мадэль",
+    "cameras_one": "{{count}} камера",
+    "cameras_few": "{{count}} камеры",
+    "cameras_many": "{{count}} камер",
+    "scene": {
+      "label": "Сцэна",
+      "description": "Асяроддзе, для якога прызначана гэтая мадэль. Камера выбірае мадэль, задаючы тую самую сцэну ў сваіх наладах дэтэктавання, а мадэль са сцэнай all выкарыстоўвае любая камера, якая нічога не задала."
+    },
+    "scenes": {
+      "all": "Усе камеры",
+      "indoor": "У памяшканні",
+      "outdoor": "На вуліцы",
+      "indoor_thermal": "У памяшканні, цеплавізійная",
+      "outdoor_thermal": "На вуліцы, цеплавізійная"
+    },
+    "hardware": {
+      "label": "Абсталяванне",
+      "placeholder": "Выберыце абсталяванне",
+      "loading": "Пошук абсталявання для дэтэктавання…",
+      "none": "Абсталяванне не выбрана",
+      "claimedBy": "выкарыстоўвае {{scene}}",
+      "detectorCount": "Дэтэктары",
+      "countRecommended_one": "Рэкамендуецца для {{count}} камеры",
+      "countRecommended_few": "Рэкамендуецца для {{count}} камер",
+      "countRecommended_many": "Рэкамендуецца для {{count}} камер",
+      "unrecognized": "Гэтая мадэль наладжана для абсталявання, якога няма ў сістэме: {{devices}}",
+      "description": "Абсталяванне, на якім гэтая мадэль выконвае дэтэктаванне.",
+      "detectorCountDescription": "Колькі працэсаў дэтэктавання запускаць на гэтым абсталяванні. Большая колькасць дэтэктараў спраўляецца з большай колькасцю камер, але займае больш памяці прылады.",
+      "unitsDescription": "Кожны блок выконвае свой працэс дэтэктавання. Блок, які ўжо выкарыстоўвае іншая мадэль, выбраць нельга."
+    },
+    "tabs": {
+      "plus": "Frigate+",
+      "custom": "Свая мадэль"
+    },
+    "plusModel": {
+      "noModelSelected": "Выберыце мадэль Frigate+"
     }
   }
 }

web/public/locales/be/views/system.json

@@ -9,7 +9,8 @@
       "go2rtc": "Журналы go2rtc - Frigate",
       "nginx": "Журналы Nginx - Frigate",
       "websocket": "Журналы паведамленняў - Frigate"
-    }
+    },
+    "health": "Стан - Frigate"
   },
   "title": "Сістэма",
   "metrics": "Сістэмныя метрыкі",
@@ -148,6 +149,10 @@
       "unused": {
         "title": "Нявыкарыстана",
         "tips": "Гэта значэнне можа недакладна адлюстроўваць вольнае месца, даступнае Frigate, калі на дыску ёсць іншыя файлы акрамя запісаў Frigate. Frigate не адсочвае месца па-за сваімі запісамі."
+      },
+      "subStream": {
+        "information": "Звесткі пра сховішча дадатковай плыні",
+        "tips": "Камеры, якія пішуць дадатковую плынь, паказваюць выкарыстанне зыходнай і нізкай якасці асобна, тымі ж назвамі, што і параметры якасці пры прайграванні запісаў. Запісы нізкай якасці застаюцца на дыску да заканчэння тэрміну захоўвання, таму гэты падзел можа паказвацца і пасля таго, як запіс дадатковай плыні быў адключаны."
       }
     }
   },
@@ -181,7 +186,11 @@
         "segmentLength": "Даўжыня сегмента запісу:",
         "ok": "Ключавыя кадры прыблізна кожныя {{seconds}} с, гэта добра для запісу і прайгравання.",
         "unknown": "Не ўдалося вызначыць інтэрвал ключавых кадраў.",
-        "recordDisabled": "Для гэтай камеры запіс адключаны."
+        "recordDisabled": "Для гэтай камеры запіс адключаны.",
+        "warningFixed": "Ключавыя кадры ідуць раўнамерна, але рэдка (прыкладна кожныя {{seconds}} с). Запіс працуе, але жывое прайграванне і перамотка пачынаюцца павольней. Задайце інтэрвал I-кадраў (ключавых кадраў) камеры роўным яе частаце кадраў.",
+        "warningVariable": "Інтэрвал ключавых кадраў нераўнамерны (ад {{minSeconds}} с да {{maxSeconds}} с), што звычайна значыць уключаны разумны кодэк (H.264+/H.265+). Так рабіць не рэкамендуецца.",
+        "errorFixed": "Ключавыя кадры кожныя прыкладна {{seconds}} с - гэта даўжэй за сегмент запісу ({{segmentTime}} с), таму некаторыя сегменты застаюцца без ключавога кадра і не будуць прайгравацца. Паменшыце інтэрвал I-кадраў (ключавых кадраў) камеры да яе частаты кадраў.",
+        "errorVariable": "Прамежкі паміж ключавымі кадрамі дасягаюць прыкладна {{seconds}} с - гэта даўжэй за сегмент запісу ({{segmentTime}} с). Некаторыя сегменты застануцца без ключавога кадра, што ламае прайграванне. Адключыце разумны кодэк (+) на камеры або паменшыце інтэрвал ключавых кадраў."
       }
     },
     "framesAndDetections": "Кадры / дэтэктаванні",
@@ -234,7 +243,11 @@
     "cameraIsOffline": "{{camera}} па-за сеткай",
     "detectIsSlow": "{{detect}} працуе павольна ({{speed}} мс)",
     "detectIsVerySlow": "{{detect}} працуе вельмі павольна ({{speed}} мс)",
-    "debugReplayActive": "Сеанс адладачнага паўтору актыўны"
+    "debugReplayActive": "Сеанс адладачнага паўтору актыўны",
+    "cameraSkippedDetections": "{{camera}} прапускае дэтэктаванне на {{pct}}% кадраў",
+    "systemNotices_one": "{{count}} сістэмная заўвага",
+    "systemNotices_other": "{{count}} сістэмнай заўвагі",
+    "retentionUnmet": "Запісы выдаляюцца да заканчэння тэрміну захоўвання, каб вызваліць месца"
   },
   "enrichments": {
     "title": "Узбагачэнні",
@@ -262,5 +275,93 @@
       "classification_speed": "Хуткасць класіфікацыі {{name}}",
       "classification_events_per_second": "Падзей класіфікацыі {{name}} на секунду"
     }
+  },
+  "health": {
+    "title": "Стан",
+    "notices": {
+      "title": "Заўвагі",
+      "empty": "Ваша ўсталёўка Frigate у парадку",
+      "dismiss": "Схаваць",
+      "openSettings": "Адкрыць налады",
+      "openLink": "Адкрыць спасылку",
+      "noMatches": "Няма заўваг, якія адпавядаюць фільтру",
+      "dismissedTitle": "Схаваныя",
+      "noneDismissed": "Няма схаваных заўваг",
+      "clearDismissed": "Ачысціць схаваныя",
+      "clearDismissedTitle": "Ачысціць схаваныя заўвагі?",
+      "clearDismissedDesc": "Усе схаваныя заўвагі будуць выдалены. Праверкі канфігурацыі і плыняў пакажуцца адразу, а іншыя заўвагі вернуцца пры наступным узнікненні.",
+      "firstSeen_one": "Упершыню {{time}}",
+      "firstSeen_other": "Упершыню {{time}} · {{count}} разу",
+      "dismissedAt": "Схавана {{time}}",
+      "filter": {
+        "showDismissed": "Паказаць схаваныя",
+        "severity": "Узровень важнасці",
+        "error": "Памылка",
+        "warning": "Папярэджанне",
+        "info": "Звесткі"
+      },
+      "kinds": {
+        "detector_stuck": "Дэтэктар {{detector}} перазапушчаны пасля таго, як перастаў адказваць",
+        "model_download_failed": "Не ўдалося спампаваць {{file}} для {{model}}: {{error}}",
+        "skipped_detections": "Дэтэктаванне не паспявала і больш за хвіліну прапускала {{pct}}% кадраў",
+        "shm_too_low": "Выдзеленую памяць /dev/shm ({{total}} МБ) трэба павялічыць прынамсі да {{min}} МБ",
+        "failed_login_one": "Няўдалая спроба ўваходу для {{user}}",
+        "failed_login_other": "Няўдалыя спробы ўваходу для {{user}}",
+        "update_available": "Даступная Frigate {{version}}"
+      },
+      "streamPrefix": "Плынь {{index}}: {{message}}",
+      "streamPrefixRestream": "Плынь {{index}} (праз go2rtc): {{message}}",
+      "streamProbeFailed": "Не ўдалося зандзіраваць плынь {{index}}: {{error}}",
+      "cameraProbeFailed": "Не ўдалося зандзіраваць плыні: {{error}}"
+    },
+    "hardware": {
+      "title": "Абсталяванне",
+      "objectDetection": "Дэтэктаванне аб'ектаў",
+      "hardwareAcceleration": "Апаратнае паскарэнне",
+      "enrichments": {
+        "title": "Узбагачэнні",
+        "semantic_search": "Семантычны пошук",
+        "face_recognition": "Распазнаванне асобы",
+        "lpr": "Распазнаванне нумарных знакаў",
+        "audio_transcription": "Транскрыпцыя гуку"
+      },
+      "cameraConnections": "Злучэнні камер",
+      "allCamerasExcellent": "Усе камеры маюць выдатнае злучэнне.",
+      "fps": "{{camera}} / {{expected}} fps",
+      "nothingConfigured": "Нічога не наладжана",
+      "allCameras": "Усе камеры",
+      "probeUnavailable": "Выяўленне абсталявання недаступнае",
+      "justStarted": "Frigate толькі што запусціўся",
+      "deviceNotFound": "{{devices}} не знойдзена ў гэтай сістэме",
+      "detectorNotRunning": "Працэс дэтэктара не запушчаны",
+      "inferenceVerySlow": "Вывядзенне вельмі павольнае ({{speed}} мс)",
+      "inferenceSlow": "Вывядзенне павольнае ({{speed}} мс)",
+      "inferenceMs": "{{speed}} мс",
+      "customArgs": "Свае аргументы ffmpeg",
+      "customArgsNotVerified": "Свае аргументы ffmpeg, не правераныя",
+      "hwaccelNotConfigured": "Апаратнае дэкадаванне не выкарыстоўваецца, хоць {{family}} даступнае. Выберыце яго ў наладах ffmpeg",
+      "hwaccelHardwareMissing": "{{family}} наладжана, але апаратнае зандзіраванне яго не выявіла",
+      "unrecognizedDevice": "Нераспазнанае перавызначэнне прылады, не праверана",
+      "decoderUsage": "дэкадавальнік {{usage}}",
+      "remoteProvider": "Аддалены правайдар",
+      "acceleratorMissing": "Наладжана для {{device}}, але сумяшчальнага паскаральніка не знойдзена",
+      "modelNotRunYet": "Мадэль яшчэ не запускалася, прылада правяраецца пасля першага выкарыстання",
+      "fellBackToCpu": "Наладжана для {{device}}, але працуе на CPU",
+      "cpuDespiteAccelerator": "Працуе на CPU, хоць паскаральнік даступны",
+      "probed": "Зандзіравана",
+      "probing": "Зандзіраванне…",
+      "recheck": "Праверыць абсталяванне зноў",
+      "cameraStreams": "Плыні камер",
+      "streamsNotChecked": "Яшчэ не правяралася",
+      "streamsChecking": "Правярае {{done}} з {{total}}",
+      "streamsClean": "Праблем з плынямі няма",
+      "checked": "Праверана",
+      "runAgain": "Запусціць зноў",
+      "runStreamChecks": "Запусціць праверку плыняў",
+      "streamsChecked_one": "{{count}} камера правераная",
+      "streamsChecked_other": "{{count}} камеры правераных",
+      "streamsFlagged_one": "{{count}} з праблемамі, пералічана ў Заўвагах",
+      "streamsFlagged_other": "{{count}} з праблемамі, пералічана ў Заўвагах"
+    }
   }
 }

web/public/locales/bg/components/player.json

@@ -1,14 +1,6 @@
 {
   "noPreviewFoundFor": "Не е намерен предварителен преглед за {{cameraName}}",
   "stats": {
-    "latency": {
-      "short": {
-        "value": "{{seconds}} сек",
-        "title": "Закъснение"
-      },
-      "title": "Закъснение:",
-      "value": "{{seconds}} секунди"
-    },
     "streamType": {
       "title": "Тип поток:",
       "short": "Тип"

web/public/locales/bs/components/player.json

@@ -21,14 +21,6 @@
       "title": "Širina pojasa:",
       "short": "Širina pojasa"
     },
-    "latency": {
-      "title": "Kasnjenje:",
-      "value": "{{seconds}} sekundi",
-      "short": {
-        "title": "Kasnjenje",
-        "value": "{{seconds}} sek"
-      }
-    },
     "totalFrames": "Ukupno okvira:",
     "droppedFrames": {
       "title": "Izgubljeni okviri:",

web/public/locales/ca/common.json

@@ -55,9 +55,9 @@
       "be": "Беларуская (belarús)"
     },
     "system": "Sistema",
-    "systemMetrics": "Mètriques del sistema",
+    "systemMetrics": "Salut i Mètriques",
     "configuration": "Configuració",
-    "systemLogs": "Registres del sistema",
+    "systemLogs": "Registres",
     "configurationEditor": "Editor de configuració",
     "languages": "Idiomes",
     "settings": "Opcions",
@@ -287,7 +287,10 @@
     "savingAll": "S'està desant tot…",
     "undoAll": "Desfés-ho tot",
     "applying": "S'està aplicant…",
-    "retry": "Torna a intentar"
+    "retry": "Torna a intentar",
+    "expand": "Expandeix",
+    "collapse": "Redueix",
+    "reload": "Torna a carregar"
   },
   "toast": {
     "copyUrlToClipboard": "URL copiada al porta-retalls.",
@@ -328,5 +331,36 @@
   "validation_errors": "Errors de validació",
   "credentialField": {
     "savedPlaceholder": "Desat — deixa en blanc per mantenir l'actual"
+  },
+  "error": {
+    "title": "Aquesta pàgina ha deixat de funcionar",
+    "desc": "La pàgina ha tingut un error inesperat. La recàrrega normalment la neteja. Si continua passant, copieu els detalls i adjunteu-los a una discussió sobre GitHub.",
+    "staleTitle": "Recàrrega requerida",
+    "staleDesc": "No s'ha pogut carregar part de la interfície, cosa que normalment significa que Frigate s'ha actualitzat mentre aquesta pestanya estava oberta. Torna a carregar per a recollir la nova versió.",
+    "partial": "Part de la interfície va deixar de funcionar.",
+    "copyDetails": "Copia els detalls",
+    "copyFailed": "No s'han pogut copiar els detalls al porta-retalls"
+  },
+  "commandMenu": {
+    "title": "Menú d'ordres",
+    "description": "Cerca pàgines, càmeres, paràmetres i accions",
+    "placeholder": "Cerca pàgines, càmeres, paràmetres i accions",
+    "empty": "Sense coincidències",
+    "hint": "{{key}} o / per obrir, Intro per executar",
+    "section": {
+      "recent": "Recent",
+      "pages": "Pàg",
+      "cameras": "Càmeres",
+      "cameraGroups": "Grups de càmera",
+      "settings": "Configuració",
+      "actions": "Accions"
+    },
+    "camera": {
+      "live": "Directe",
+      "review": "Revisió"
+    },
+    "action": {
+      "toggleTheme": "Commuta el mode clar i fosc"
+    }
   }
 }

web/public/locales/ca/components/dialog.json

@@ -137,6 +137,15 @@
         "partial": "S'han iniciat {{successful}} de {{total}} exportacions. Ha fallat: {{failedItems}}",
         "failed": "No s'han pogut iniciar {{total}} exportacions. Ha fallat: {{failedItems}}"
       }
+    },
+    "stream": {
+      "label": "Qualitat",
+      "auto": "Automàtic",
+      "main": "Original",
+      "sub": "Baix",
+      "autoDesc": "Utilitza el flux principal i torna al flux de menor qualitat quan el principal no està disponible.",
+      "mainDesc": "Exporta només el flux principal de qualitat original. Faltaran els intervals de temps on el flux principal no està disponible.",
+      "subDesc": "Exporta només el subflux de menor qualitat. Faltaran els intervals de temps on el subflux no està disponible."
     }
   },
   "streaming": {
@@ -172,7 +181,8 @@
       "deleteNow": "Suprimir ara",
       "export": "Exportar",
       "markAsReviewed": "Marcar com a revisat",
-      "markAsUnreviewed": "Marcar com no revisat"
+      "markAsUnreviewed": "Marcar com no revisat",
+      "generateDescription": "Genera la descripció"
     },
     "confirmDelete": {
       "title": "Confirmar la supressió",
@@ -191,6 +201,12 @@
       "custom": "Marca horària personalitzada",
       "button": "Comparteix l'URL de la marca horària",
       "shareTitle": "Marca de temps de revisió de Frigate: {{camera}}"
+    },
+    "genaiDescription": {
+      "toast": {
+        "success": "S'ha sol·licitat una descripció de IA Generativa per a aquest element de revisió.",
+        "error": "No s'ha pogut sol·licitar la descripció: {{error}}"
+      }
     }
   },
   "imagePicker": {

web/public/locales/ca/components/player.json

@@ -1,13 +1,5 @@
 {
   "stats": {
-    "latency": {
-      "short": {
-        "title": "Latència",
-        "value": "{{seconds}} s"
-      },
-      "title": "Latència:",
-      "value": "{{seconds}} segons"
-    },
     "streamType": {
       "title": "Tipus de transmissió:",
       "short": "Tipus"
@@ -45,8 +37,23 @@
       "submittedFrigatePlus": "Fotograma enviat correctament a Frigate+"
     },
     "error": {
-      "submitFrigatePlusFailed": "Error al enviar fotograma a Frigate+"
+      "submitFrigatePlusFailed": "Error al enviar fotograma a Frigate+",
+      "playRecordingsFailed": "No s'han pogut reproduir els enregistraments (error {{code}}): {{message}}"
     }
   },
-  "cameraOff": "La càmera està apagada"
+  "cameraOff": "La càmera està apagada",
+  "quality": {
+    "auto": "Automàtic",
+    "autoLow": "Reproduint de baixa qualitat (ample de banda limitat)",
+    "autoLowCodec": "Reproducció de baixa qualitat (Original no suportat per aquest navegador)",
+    "autoLowSaveData": "Reproducció de baixa qualitat (desador de dades)",
+    "main": "Original",
+    "sub": "Baixa",
+    "notSupportedBrowser": "No és compatible amb aquest navegador",
+    "label": "Qualitat",
+    "noAudio": "Sense àudio",
+    "audioRate": "{{rate}} àudio kHz",
+    "audioCodecRate": "{{codec}} {{rate}} kHz",
+    "noRecordings": "No hi ha enregistraments en aquest interval de temps"
+  }
 }

web/public/locales/ca/config/cameras.json

@@ -46,6 +46,10 @@
         "num_threads": {
             "label": "Fils de detecció",
             "description": "Nombre de fils a utilitzar per al processament de detecció d'àudio."
+        },
+        "labelmap": {
+            "label": "Personalització del mapa d'etiquetes d'àudio",
+            "description": "Sobreescriu o reassigna les entrades per a fusionar-se en el mapa d'etiquetes d'àudio estàndard."
         }
     },
     "audio_transcription": {
@@ -73,6 +77,10 @@
         "order": {
             "label": "Posició",
             "description": "Posició numèrica que controla l'ordenació de la càmera en la disposició Birdseye."
+        },
+        "modes": {
+            "label": "Tipus d'activitat",
+            "description": "Tipus d'activitat que inclouen càmeres a Birdseye."
         }
     },
     "detect": {
@@ -133,6 +141,10 @@
         "annotation_offset": {
             "label": "Desplaçament de l'anotació",
             "description": "Mil·lisegons per a desplaçar detecta anotacions per a alinear millor els límits de la línia de temps amb els enregistraments; pot ser positiu o negatiu."
+        },
+        "scene": {
+            "label": "Detecta l'escena",
+            "description": "L'entorn que mira aquesta càmera, utilitzat per triar quin dels models configurats s'executa en ell. Les càmeres que queden a 'totes' executen el model configurat amb una escena de 'totes'."
         }
     },
     "face_recognition": {
@@ -176,6 +188,10 @@
             "record": {
                 "label": "Registra els arguments de sortida",
                 "description": "Arguments de sortida predeterminats per a enregistrar fluxos de rols."
+            },
+            "record_sub": {
+                "label": "Arguments de sortida del registre de subflux",
+                "description": "Arguments de sortida per als fluxos de rols record.sub. Els arguments de sortida del registre s'utilitzen quan no s'estableix."
             }
         },
         "retry_interval": {
@@ -356,51 +372,51 @@
             "label": "Màscara en brut"
         },
         "genai": {
-            "label": "Configuració de l'objecte GenAI",
-            "description": "Opcions de GenAI per descriure objectes rastrejats i enviar fotogrames per a la generació.",
+            "label": "Configuració de l'objecte IA Generativa",
+            "description": "Opcions d'IA Generativa per descriure objectes rastrejats i enviar fotogrames per a la generació.",
             "enabled": {
-                "label": "Habilita el GenAI",
-                "description": "Habilita la generació de descripcions de GenAI per als objectes rastrejats de manera predeterminada."
+                "label": "Habilita l'IA Generativa",
+                "description": "Habilita la generació de descripcions d'IA Generativa per als objectes rastrejats de manera predeterminada."
             },
             "use_snapshot": {
                 "label": "Utilitza instantànies",
-                "description": "Usa instantànies d'objecte en lloc de miniatures per a la generació de descripcions de GenAI."
+                "description": "Usa instantànies d'objecte en lloc de miniatures per a la generació de descripcions d'IA Generativa."
             },
             "prompt": {
                 "label": "Indicació de la llegenda",
-                "description": "Plantilla de pregunta predeterminada utilitzada en generar descripcions amb GenAI."
+                "description": "Plantilla de pregunta predeterminada utilitzada en generar descripcions amb IA Generativa."
             },
             "object_prompts": {
                 "label": "Peticions d'objecte",
-                "description": "Per objecte demana personalitzar les sortides de GenAI per a etiquetes específiques."
+                "description": "Per objecte demana personalitzar les sortides d'IA Generativa per a etiquetes específiques."
             },
             "objects": {
-                "label": "Objectes GenAI",
-                "description": "Llista d'etiquetes d'objectes a enviar a GenAI per defecte."
+                "label": "Objectes d'IA Generativa",
+                "description": "Llista d'etiquetes d'objectes a enviar a IA Generativa per defecte."
             },
             "required_zones": {
                 "label": "Zones requerides",
-                "description": "Zones que s'han d'introduir perquè els objectes es puguin classificar per a la generació de descripcions de GenAI."
+                "description": "Zones que s'han d'introduir perquè els objectes es puguin classificar per a la generació de descripcions d'IA Generativa."
             },
             "debug_save_thumbnails": {
                 "label": "Desa les miniatures",
-                "description": "Desa les miniatures enviades a GenAI per a la depuració i la revisió."
+                "description": "Desa les miniatures enviades a IA Generativa per a la depuració i la revisió."
             },
             "send_triggers": {
-                "label": "Activadors de GenAI",
-                "description": "Defineix quan s'han d'enviar fotogrames a GenAI (al final, després de les actualitzacions, etc.).",
+                "label": "Activadors d'IA Generativa",
+                "description": "Defineix quan s'han d'enviar fotogrames a IA Generativa (al final, després de les actualitzacions, etc.).",
                 "tracked_object_end": {
                     "label": "Envia al final",
-                    "description": "Envia una sol·licitud a GenAI quan acabi l'objecte rastrejat."
+                    "description": "Envia una sol·licitud a IA Generativa quan acabi l'objecte rastrejat."
                 },
                 "after_significant_updates": {
-                    "label": "Activador de GenAI primerenc",
-                    "description": "Envia una sol·licitud a GenAI després d'un nombre especificat d'actualitzacions significatives per a l'objecte rastrejat."
+                    "label": "Activador d'IA Generativa primerenc",
+                    "description": "Envia una sol·licitud a IA Generativa després d'un nombre especificat d'actualitzacions significatives per a l'objecte rastrejat."
                 }
             },
             "enabled_in_config": {
-                "label": "Estat original de GenAI",
-                "description": "Indica si el GenAI s'ha activat a la configuració estàtica original."
+                "label": "Estat original d'IA Generativa",
+                "description": "Indica si l'IA Generativa s'ha activat a la configuració estàtica original."
             }
         }
     },
@@ -505,11 +521,59 @@
         "enabled_in_config": {
             "label": "Estat de l'enregistrament original",
             "description": "Indica si l'enregistrament s'ha activat en la configuració estàtica original."
+        },
+        "sub": {
+            "label": "Enregistrament de subflux",
+            "description": "Paràmetres per a enregistrar un segon flux de menor qualitat.",
+            "enabled": {
+                "label": "Habilita l'enregistrament de subflux",
+                "description": "Permet l'enregistrament d'un segon flux de menor qualitat per a la reproducció de qualitat adaptativa i la retenció estesa."
+            },
+            "continuous": {
+                "label": "Retenció contínua del subflux",
+                "description": "Nombre de dies per a retenir els enregistraments de sub flux independentment dels objectes o el moviment seguits.",
+                "days": {
+                    "label": "Dies de retenció",
+                    "description": "Dies per retenir enregistraments."
+                }
+            },
+            "motion": {
+                "label": "Retenció del moviment del subflux",
+                "description": "Nombre de dies per a retenir els enregistraments de substream activats pel moviment.",
+                "days": {
+                    "label": "Dies de retenció",
+                    "description": "Dies per retenir enregistraments."
+                }
+            },
+            "alerts": {
+                "label": "Retenció d'alertes del subflux",
+                "description": "Paràmetres de retenció per als enregistraments de substream d'alertes.",
+                "days": {
+                    "label": "Dies de retenció",
+                    "description": "Nombre de dies per a retenir enregistraments d'esdeveniments de detecció."
+                },
+                "mode": {
+                    "label": "Mode de retenció",
+                    "description": "Mode de retenció: tot (desa tots els segments), moviment (desa els segments amb moviment), o actiuobobjectes (desa els segments amb objectes actius)."
+                }
+            },
+            "detections": {
+                "label": "Retenció de la detecció de subflux",
+                "description": "Paràmetres de retenció per als enregistraments de substream de les deteccions.",
+                "days": {
+                    "label": "Dies de retenció",
+                    "description": "Nombre de dies per a retenir enregistraments d'esdeveniments de detecció."
+                },
+                "mode": {
+                    "label": "Mode de retenció",
+                    "description": "Mode de retenció: tot (desa tots els segments), moviment (desa els segments amb moviment), o actiuobobjectes (desa els segments amb objectes actius)."
+                }
+            }
         }
     },
     "review": {
         "label": "Revisió",
-        "description": "Configuració que controla les alertes, les deteccions i els resums de revisió de GenAI utilitzats per la interfície d'usuari i l'emmagatzematge d'aquesta càmera.",
+        "description": "Configuració que controla les alertes, les deteccions i els resums de revisió d'IA Generativa utilitzats per la interfície d'usuari i l'emmagatzematge d'aquesta càmera.",
         "alerts": {
             "label": "Configuració d'alertes",
             "description": "Paràmetres per als quals els objectes rastrejats generen alertes i com es mantenen les alertes",
@@ -559,43 +623,51 @@
             }
         },
         "genai": {
-            "label": "Configuració del GenAI",
-            "description": "Controla l'ús de la IA generativa per a la producció de descripcions i resums d'elements de revisió.",
+            "label": "Configuració de l'IA Generativa",
+            "description": "Controla l'ús de l'IA generativa per a la producció de descripcions i resums d'elements de revisió.",
             "enabled": {
-                "label": "Habilita les descripcions del GenAI",
-                "description": "Activa o desactiva les descripcions i resums generats per GenAI per als elements de revisió."
+                "label": "Habilita les descripcions de l'IA Generativa",
+                "description": "Activa o desactiva les descripcions i resums generats per IA Generativa per als elements de revisió."
             },
             "alerts": {
-                "label": "Habilita el GenAI per a alertes",
-                "description": "Utilitzeu GenAI per a generar descripcions per als elements d'alerta."
+                "label": "Habilita l'IA Generativa per a alertes",
+                "description": "Utilitzeu IA Generativa per a generar descripcions per als elements d'alerta."
             },
             "detections": {
-                "label": "Habilita el GenAI per a les deteccions",
-                "description": "Utilitzeu GenAI per generar descripcions per als elements de detecció."
+                "label": "Habilita l'IA Generativa per a les deteccions",
+                "description": "Utilitzeu IA Generativa per generar descripcions per als elements de detecció."
             },
             "image_source": {
                 "label": "Revisa l'origen de la imatge",
-                "description": "Font d'imatges enviades a GenAI ('previsualització' o 'enregistraments'); 'enregistraments' utilitza Fotogrames de més qualitat però més tokens."
+                "description": "Font d'imatges enviades a IA Generativa ('previsualització' o 'enregistraments'); 'enregistraments' utilitza Fotogrames de més qualitat però més tokens."
             },
             "additional_concerns": {
                 "label": "Altres preocupacions",
-                "description": "Una llista de preocupacions o notes addicionals que el GenAI ha de tenir en compte a l'hora d'avaluar l'activitat en aquesta càmera."
+                "description": "Una llista de preocupacions o notes addicionals que l¡IA Generativa ha de tenir en compte a l'hora d'avaluar l'activitat en aquesta càmera."
             },
             "debug_save_thumbnails": {
                 "label": "Desa les miniatures",
-                "description": "Desa les miniatures que s'envien al proveïdor GenAI per a la depuració i la revisió."
+                "description": "Desa les miniatures que s'envien al proveïdor IA Generativa per a la depuració i la revisió."
             },
             "enabled_in_config": {
-                "label": "Estat original de GenAI",
-                "description": "Fa un seguiment de si la revisió de GenAI es va habilitar originalment a la configuració estàtica."
+                "label": "Estat original d'IA Generativa",
+                "description": "Fa un seguiment de si la revisió d'IA Generativa es va habilitar originalment a la configuració estàtica."
             },
             "preferred_language": {
                 "label": "Idioma preferit",
-                "description": "Idioma preferit per sol·licitar al proveïdor GenAI respostes generades."
+                "description": "Idioma preferit per sol·licitar al proveïdor d'IA Generativa respostes generades."
             },
             "activity_context_prompt": {
                 "label": "Indicador de context de l'activitat",
                 "description": "Pregunta personalitzada que descriu el que és i no és una activitat sospitosa per proporcionar context per als resums de GenAI."
+            },
+            "response_style": {
+                "label": "Estil de resposta",
+                "description": "Escriviu un predefinit d'estil per a les descripcions generades de les revisions. Els valors predefinits ajusten el to i el nivell de detall del títol orientat a l'usuari, el resum i la descripció de l'escena; «predeterminat» deixa l'indicatiu incorporat sense canvis."
+            },
+            "frame_mode": {
+                "label": "Mode de fotograma",
+                "description": "Com es presenten els marcs al model. Els 'frames' envien l'indicatiu seguit pels fotogrames, que s'adapta als models que segueixen una seqüència per si sols. 'annotated.frames' etiqueta cada marc i entrellaça notes derivades del seguiment d'objectes, el que ajuda als models que perden el rastre de l'activitat que es repeteix o reverteix."
             }
         }
     },

web/public/locales/ca/config/global.json

@@ -31,6 +31,10 @@
             "record": {
                 "label": "Registra els arguments de sortida",
                 "description": "Arguments de sortida predeterminats per a enregistrar fluxos de rols."
+            },
+            "record_sub": {
+                "label": "Arguments de sortida del registre de subflux",
+                "description": "Arguments de sortida per als fluxos de rols record.sub. Els arguments de sortida del registre s'utilitzen quan no s'estableix."
             }
         },
         "retry_interval": {
@@ -193,49 +197,49 @@
         "genai": {
             "required_zones": {
                 "label": "Zones requerides",
-                "description": "Zones que s'han d'introduir perquè els objectes es puguin classificar per a la generació de descripcions de GenAI."
+                "description": "Zones que s'han d'introduir perquè els objectes es puguin classificar per a la generació de descripcions d'IA Generativa."
             },
-            "label": "Configuració de l'objecte GenAI",
-            "description": "Opcions de GenAI per descriure objectes rastrejats i enviar fotogrames per a la generació.",
+            "label": "Configuració de l'objecte IA Generativa",
+            "description": "Opcions d'IA Generativa per descriure objectes rastrejats i enviar fotogrames per a la generació.",
             "enabled": {
-                "label": "Habilita el GenAI",
-                "description": "Habilita la generació de descripcions de GenAI per als objectes rastrejats de manera predeterminada."
+                "label": "Habilita l'IA Generativa",
+                "description": "Habilita la generació de descripcions d'IA Generativa per als objectes rastrejats de manera predeterminada."
             },
             "use_snapshot": {
                 "label": "Utilitza instantànies",
-                "description": "Usa instantànies d'objecte en lloc de miniatures per a la generació de descripcions de GenAI."
+                "description": "Usa instantànies d'objecte en lloc de miniatures per a la generació de descripcions d'IA Generativa."
             },
             "prompt": {
                 "label": "Indicació de la llegenda",
-                "description": "Plantilla de pregunta predeterminada utilitzada en generar descripcions amb GenAI."
+                "description": "Plantilla de pregunta predeterminada utilitzada en generar descripcions amb IA Generativa."
             },
             "object_prompts": {
                 "label": "Peticions d'objecte",
-                "description": "Per objecte demana personalitzar les sortides de GenAI per a etiquetes específiques."
+                "description": "Per objecte demana personalitzar les sortides d'IA Generativa per a etiquetes específiques."
             },
             "objects": {
-                "label": "Objectes GenAI",
-                "description": "Llista d'etiquetes d'objectes a enviar a GenAI per defecte."
+                "label": "Objectes d'IA Generativa",
+                "description": "Llista d'etiquetes d'objectes a enviar a IA Generativa per defecte."
             },
             "debug_save_thumbnails": {
                 "label": "Desa les miniatures",
-                "description": "Desa les miniatures enviades a GenAI per a la depuració i la revisió."
+                "description": "Desa les miniatures enviades a IA Generativa per a la depuració i la revisió."
             },
             "send_triggers": {
-                "label": "Activadors de GenAI",
-                "description": "Defineix quan s'han d'enviar fotogrames a GenAI (al final, després de les actualitzacions, etc.).",
+                "label": "Activadors d'IA Generativa",
+                "description": "Defineix quan s'han d'enviar fotogrames a IA Generativa (al final, després de les actualitzacions, etc.).",
                 "tracked_object_end": {
                     "label": "Envia al final",
-                    "description": "Envia una sol·licitud a GenAI quan acabi l'objecte rastrejat."
+                    "description": "Envia una sol·licitud a IA Generativa quan acabi l'objecte rastrejat."
                 },
                 "after_significant_updates": {
-                    "label": "Activador de GenAI primerenc",
-                    "description": "Envia una sol·licitud a GenAI després d'un nombre especificat d'actualitzacions significatives per a l'objecte rastrejat."
+                    "label": "Activador d'IA Generativa primerenc",
+                    "description": "Envia una sol·licitud a IA Generativa després d'un nombre especificat d'actualitzacions significatives per a l'objecte rastrejat."
                 }
             },
             "enabled_in_config": {
-                "label": "Estat original de GenAI",
-                "description": "Indica si el GenAI s'ha activat a la configuració estàtica original."
+                "label": "Estat original d'IA Generativa",
+                "description": "Indica si l'IA Generativa s'ha activat a la configuració estàtica original."
             }
         },
         "label": "Objectes",
@@ -388,6 +392,54 @@
         "enabled_in_config": {
             "label": "Estat de l'enregistrament original",
             "description": "Indica si l'enregistrament s'ha activat en la configuració estàtica original."
+        },
+        "sub": {
+            "label": "Enregistrament de subflux",
+            "description": "Paràmetres per a enregistrar un segon flux de menor qualitat.",
+            "enabled": {
+                "label": "Habilita l'enregistrament de subflux",
+                "description": "Permet l'enregistrament d'un segon flux de menor qualitat per a la reproducció de qualitat adaptativa i la retenció estesa."
+            },
+            "continuous": {
+                "label": "Retenció contínua del subflux",
+                "description": "Nombre de dies per a retenir els enregistraments de sub flux independentment dels objectes o el moviment seguits.",
+                "days": {
+                    "label": "Dies de retenció",
+                    "description": "Dies per retenir enregistraments."
+                }
+            },
+            "motion": {
+                "label": "Retenció del moviment del subflux",
+                "description": "Nombre de dies per a retenir els enregistraments de substream activats pel moviment.",
+                "days": {
+                    "label": "Dies de retenció",
+                    "description": "Dies per retenir enregistraments."
+                }
+            },
+            "alerts": {
+                "label": "Retenció d'alertes del subflux",
+                "description": "Paràmetres de retenció per als enregistraments de substream d'alertes.",
+                "days": {
+                    "label": "Dies de retenció",
+                    "description": "Nombre de dies per a retenir enregistraments d'esdeveniments de detecció."
+                },
+                "mode": {
+                    "label": "Mode de retenció",
+                    "description": "Mode de retenció: tot (desa tots els segments), moviment (desa els segments amb moviment), o actiuobobjectes (desa els segments amb objectes actius)."
+                }
+            },
+            "detections": {
+                "label": "Retenció de la detecció de subflux",
+                "description": "Paràmetres de retenció per als enregistraments de substream de les deteccions.",
+                "days": {
+                    "label": "Dies de retenció",
+                    "description": "Nombre de dies per a retenir enregistraments d'esdeveniments de detecció."
+                },
+                "mode": {
+                    "label": "Mode de retenció",
+                    "description": "Mode de retenció: tot (desa tots els segments), moviment (desa els segments amb moviment), o actiuobobjectes (desa els segments amb objectes actius)."
+                }
+            }
         }
     },
     "review": {
@@ -417,42 +469,50 @@
         },
         "genai": {
             "image_source": {
-                "description": "Font d'imatges enviades a GenAI ('previsualització' o 'enregistraments'); 'enregistraments' utilitza Fotogrames de més qualitat però més tokens.",
+                "description": "Font d'imatges enviades a IA Generativa ('previsualització' o 'enregistraments'); 'enregistraments' utilitza Fotogrames de més qualitat però més tokens.",
                 "label": "Revisa l'origen de la imatge"
             },
             "enabled_in_config": {
-                "description": "Fa un seguiment de si la revisió de GenAI es va habilitar originalment a la configuració estàtica.",
-                "label": "Estat original de GenAI"
+                "description": "Fa un seguiment de si la revisió d'IA Generativa es va habilitar originalment a la configuració estàtica.",
+                "label": "Estat original d'IA Generativa"
             },
-            "label": "Configuració del GenAI",
-            "description": "Controla l'ús de la IA generativa per a la producció de descripcions i resums d'elements de revisió.",
+            "label": "Configuració de l'IA Generativa",
+            "description": "Controla l'ús de l'IA generativa per a la producció de descripcions i resums d'elements de revisió.",
             "enabled": {
-                "label": "Habilita les descripcions del GenAI",
-                "description": "Activa o desactiva les descripcions i resums generats per GenAI per als elements de revisió."
+                "label": "Habilita les descripcions de l'IA Generativa",
+                "description": "Activa o desactiva les descripcions i resums generats per IA Generativa per als elements de revisió."
             },
             "alerts": {
-                "label": "Habilita el GenAI per a alertes",
-                "description": "Utilitzeu GenAI per a generar descripcions per als elements d'alerta."
+                "label": "Habilita l'IA Generativa per a alertes",
+                "description": "Utilitzeu IA Generativa per a generar descripcions per als elements d'alerta."
             },
             "detections": {
-                "label": "Habilita el GenAI per a les deteccions",
-                "description": "Utilitzeu GenAI per generar descripcions per als elements de detecció."
+                "label": "Habilita l'IA Generativa per a les deteccions",
+                "description": "Utilitzeu IA Generativa per generar descripcions per als elements de detecció."
             },
             "additional_concerns": {
                 "label": "Altres preocupacions",
-                "description": "Una llista de preocupacions o notes addicionals que el GenAI ha de tenir en compte a l'hora d'avaluar l'activitat en aquesta càmera."
+                "description": "Una llista de preocupacions o notes addicionals que l¡IA Generativa ha de tenir en compte a l'hora d'avaluar l'activitat en aquesta càmera."
             },
             "debug_save_thumbnails": {
                 "label": "Desa les miniatures",
-                "description": "Desa les miniatures que s'envien al proveïdor GenAI per a la depuració i la revisió."
+                "description": "Desa les miniatures que s'envien al proveïdor IA Generativa per a la depuració i la revisió."
             },
             "preferred_language": {
                 "label": "Idioma preferit",
-                "description": "Idioma preferit per sol·licitar al proveïdor GenAI respostes generades."
+                "description": "Idioma preferit per sol·licitar al proveïdor d'IA Generativa respostes generades."
             },
             "activity_context_prompt": {
                 "label": "Indicador de context de l'activitat",
                 "description": "Pregunta personalitzada que descriu el que és i no és una activitat sospitosa per proporcionar context per als resums de GenAI."
+            },
+            "response_style": {
+                "label": "Estil de resposta",
+                "description": "Escriviu un predefinit d'estil per a les descripcions generades de les revisions. Els valors predefinits ajusten el to i el nivell de detall del títol orientat a l'usuari, el resum i la descripció de l'escena; «predeterminat» deixa l'indicatiu incorporat sense canvis."
+            },
+            "frame_mode": {
+                "label": "Mode de fotograma",
+                "description": "Com es presenten els marcs al model. Els 'frames' envien l'indicatiu seguit pels fotogrames, que s'adapta als models que segueixen una seqüència per si sols. 'annotated.frames' etiqueta cada marc i entrellaça notes derivades del seguiment d'objectes, el que ajuda als models que perden el rastre de l'activitat que es repeteix o reverteix."
             }
         },
         "label": "Revisió",
@@ -699,7 +759,7 @@
         "description": "Configuració per a la transcripció d'àudio en viu i de veu utilitzada per a esdeveniments i llegendes en directe.",
         "language": {
             "label": "Idioma de transcripció",
-            "description": "Codi d'idioma utilitzat per a la transcripció/traducció (per exemple 'en' per a l'anglès). Vegeu https://whisper-api.com/docs/languages/ per als codis de llengua compatibles."
+            "description": "Codi d'idioma utilitzat per a la transcripció/traducció (per exemple 'en' per a l'anglès), o 'auto' per permetre que el model el detecti. Vegeu https://whisper-api.com/docs/languages/ per als codis de llengua compatibles."
         },
         "device": {
             "label": "Dispositiu de transcripció",
@@ -712,6 +772,10 @@
         "live_enabled": {
             "label": "Transcripció en viu",
             "description": "Habilita la transcripció en directe per a l'àudio a mesura que es rep."
+        },
+        "model": {
+            "label": "Model de transcripció d'àudio o nom del proveïdor GenAI",
+            "description": "El dorsal de transcripció: 'whisper' per als models locals integrats de Frigate, o el nom d'un proveïdor de GenAI amb el rol de transcriure."
         }
     },
     "version": {
@@ -971,7 +1035,7 @@
         },
         "default_role": {
             "label": "Rol predeterminat",
-            "description": "Rol predeterminat assignat als usuaris intermediaris autenticats quan no s'aplica cap mapatge de rols."
+            "description": "Rol predeterminat assignat als usuaris intermediaris autenticats quan no s'aplica cap mapatge de rols. Establiu-ho a «none» per a denegar l'accés als usuaris sense mapar."
         },
         "separator": {
             "label": "Caràcter separador",
@@ -1019,7 +1083,7 @@
         }
     },
     "genai": {
-        "label": "Configuració de la IA generada",
+        "label": "Configuració de l'IA Generativa",
         "description": "Paràmetres per als proveïdors integrats generatius d'IA utilitzats per generar descripcions d'objectes i resums de revisions.",
         "api_key": {
             "label": "Clau API",
@@ -1035,15 +1099,15 @@
         },
         "provider": {
             "label": "Proveïdor",
-            "description": "El proveïdor GenAI a utilitzar (per exemple: ollama, gemini, openai)."
+            "description": "El proveïdor IA Generativa a utilitzar (per exemple: ollama, gemini, openai)."
         },
         "roles": {
             "label": "Rols",
-            "description": "Rols de GenAI (xat, descripcions, incrustacions); un proveïdor per rol."
+            "description": "Rols d'IA Generativa (xat, descripcions, incrustacions, transcriure); un proveïdor per rol. Per defecte només es concedeixen xats, descripcions i incrustacions; la transcripció s'ha de llistar explícitament."
         },
         "provider_options": {
             "label": "Opcions del proveïdor",
-            "description": "Opcions addicionals específiques del proveïdor per passar al client GenAI."
+            "description": "Opcions addicionals específiques del proveïdor per passar al client d'IA Generativa."
         },
         "runtime_options": {
             "label": "Opcions de temps d'execució",
@@ -1084,6 +1148,10 @@
         "num_threads": {
             "label": "Fils de detecció",
             "description": "Nombre de fils a utilitzar per al processament de detecció d'àudio."
+        },
+        "labelmap": {
+            "label": "Personalització del mapa d'etiquetes d'àudio",
+            "description": "Sobreescriu o reassigna les entrades per a fusionar-se en el mapa d'etiquetes d'àudio estàndard."
         }
     },
     "birdseye": {
@@ -1132,7 +1200,11 @@
             "label": "Posició",
             "description": "Posició numèrica que controla l'ordenació de la càmera en la disposició Birdseye."
         },
-        "label": "Birdseye"
+        "label": "Birdseye",
+        "modes": {
+            "label": "Tipus d'activitat",
+            "description": "Tipus d'activitat que inclouen càmeres a Birdseye."
+        }
     },
     "detect": {
         "label": "Detecció d'objectes",
@@ -1192,6 +1264,10 @@
         "annotation_offset": {
             "label": "Desplaçament de l'anotació",
             "description": "Mil·lisegons per a desplaçar detecta anotacions per a alinear millor els límits de la línia de temps amb els enregistraments; pot ser positiu o negatiu."
+        },
+        "scene": {
+            "label": "Detecta l'escena",
+            "description": "L'entorn que mira aquesta càmera, utilitzat per triar quin dels models configurats s'executa en ell. Les càmeres que queden a 'totes' executen el model configurat amb una escena de 'totes'."
         }
     },
     "classification": {
@@ -1399,5 +1475,57 @@
     "active_profile": {
         "label": "Perfil actiu",
         "description": "Nom de perfil actualment actiu. Només en temps d'execució, no ha persistit en YAML."
+    },
+    "models": {
+        "label": "Models de detecció",
+        "description": "Models de detecció d'objectes i el maquinari en què s'executa. Les càmeres trien un model fent coincidir el seu detect.scene amb l'escena d'un model.",
+        "scene": {
+            "label": "Escena del model",
+            "description": "L'entorn de la càmera per al qual s'utilitza aquest model. Les càmeres seleccionen un model establint detect.scene a un valor coincident, i «all» és utilitzat per qualsevol càmera que no n'estableixi un."
+        },
+        "devices": {
+            "label": "hardware de detecció",
+            "description": "Aquest model funciona com a '<detector>' o '<detector>:<device>' (per exemple 'edgetpu:pci:0' o 'openvino:GPU'). La llista del mateix dispositiu més d'una vegada executa processos d'inferència addicionals."
+        },
+        "path": {
+            "label": "Ruta del model de detector d'objectes personalitzat",
+            "description": "Ruta a un fitxer de model de detecció personalitzat (o plus://<model_id> per a models Frigate+)."
+        },
+        "labelmap_path": {
+            "label": "Mapa d'etiquetes per al detector d'objectes personalitzat",
+            "description": "Ruta a un fitxer de mapa d'etiquetes que assigna classes numèriques a etiquetes de cadena per al detector."
+        },
+        "width": {
+            "label": "Amplada d'entrada del model de detecció d'objectes",
+            "description": "Amplada del tensor d'entrada del model en píxels."
+        },
+        "height": {
+            "label": "Alçada d'entrada del model de detecció d'objectes",
+            "description": "Alçada del tensor d'entrada del model en píxels."
+        },
+        "labelmap": {
+            "label": "Personalització del mapa d'etiquetes",
+            "description": "Sobreescriu o reassigna les entrades per a fusionar-se en el mapa d'etiquetes estàndard."
+        },
+        "attributes_map": {
+            "label": "Mapa d'etiquetes d'objectes a les seves etiquetes d'atribut",
+            "description": "Assignació des d'etiquetes d'objectes a etiquetes d'atribut utilitzades per adjuntar metadades (per exemple 'cotxe' -). ['matricula'])."
+        },
+        "input_tensor": {
+            "label": "Forma del sensor d'entrada del model",
+            "description": "Format del sensor esperat pel model: 'nhwc' o 'nchw'."
+        },
+        "input_pixel_format": {
+            "label": "Format de color del píxel d'entrada del model",
+            "description": "Espai de color del píxel esperat pel model: 'rgb', 'bgr' o 'yuv'."
+        },
+        "input_dtype": {
+            "label": "Tipus D d'entrada del model",
+            "description": "Tipus de dades del tensor d'entrada del model (per exemple «float32»)."
+        },
+        "model_type": {
+            "label": "Tipus de model de detecció d'objectes",
+            "description": "Tipus d'arquitectura del model de detector (ssd, yolox, yolonas, yolo-generic, rfdetr, dfine) utilitzat per alguns detectors per a l'optimització."
+        }
     }
 }

web/public/locales/ca/config/validation.json

@@ -31,5 +31,8 @@
     },
     "detect": {
         "dimensionMustBeEven": "Ha de ser un nombre parell."
+    },
+    "models": {
+        "defaultRequired": "Un model ha d'utilitzar una escena de \"Totes les càmeres\". Sense ella, qualsevol càmera que no esculli una escena no té cap model al qual tornar."
     }
 }

web/public/locales/ca/views/chat.json

@@ -55,6 +55,12 @@
         "auto_scroll": {
             "title": "Desplaçament automàtic",
             "desc": "Segueix els missatges nous a mesura que arriben."
+        },
+        "always_allow": {
+            "title": "Accions sempre permeses",
+            "desc": "Accions que l'assistent pot executar sense preguntar primer.",
+            "none": "Cap",
+            "reset": "Restableix"
         }
     },
     "stats": {
@@ -68,5 +74,15 @@
     },
     "thinking": {
         "toggle": "Commuta el pensament"
+    },
+    "approval": {
+        "title": "Aprova {{tool}}?",
+        "desc": "Aquesta acció canvia alguna cosa a Frigate. Revisa els detalls abans d'aprovar.",
+        "approve": "Aprova",
+        "always_allow": "Permet sempre",
+        "reject": "Rebutja",
+        "approved": "Aprovat",
+        "rejected": "Rebutjat",
+        "placeholder": "Aprova o rebutja l'acció pendent de continuar"
     }
 }

web/public/locales/ca/views/events.json

@@ -91,5 +91,6 @@
     "crop": "Escapça per filtrar",
     "cropAria": "Commuta la vista prèvia d'escapçament a les àrees filtrades",
     "cropDesc": "Amplia les vistes prèvies a les àrees de filtre seleccionades en lloc de mostrar el fotograma complet."
-  }
+  },
+  "subOnlyQuality": "Només els enregistraments de baixa qualitat estan disponibles en aquest interval de temps"
 }

web/public/locales/ca/views/exports.json

@@ -63,7 +63,8 @@
     "copying": "Copiant",
     "encoding": "Codificant",
     "encodingRetry": "Codificant (reintent)",
-    "finalizing": "Finalitzant"
+    "finalizing": "Finalitzant",
+    "merging": "Fusió"
   },
   "caseView": {
     "noDescription": "Sense descripció",

web/public/locales/ca/views/live.json

@@ -50,7 +50,12 @@
   "lowBandwidthMode": "Mode de baix ample de banda",
   "twoWayTalk": {
     "enable": "Activa la comunicació bidireccional",
-    "disable": "Desactiva la comunicació bidireccional"
+    "disable": "Desactiva la comunicació bidireccional",
+    "requiresWebRTC": "La conversa en dues direccions requereix WebRTC, que no està disponible",
+    "error": {
+      "microphone": "La conversa en dues direccions no ha pogut accedir al micròfon",
+      "refused": "No s'ha pogut iniciar la conversa en dues direccions. Consulteu la consola del navegador per a més detalls."
+    }
   },
   "cameraAudio": {
     "enable": "Habilitar l'àudio de la càmera",
@@ -127,14 +132,40 @@
     },
     "lowBandwidth": {
       "resetStream": "Restablir transmissió",
-      "tips": "La vista en directe està en mode de baix ample de banda a causa d'errors de transmissió o de buffering."
+      "tips": "La vista en directe està en mode de baix ample de banda a causa d'errors de transmissió o de buffering.",
+      "force": {
+        "label": "Força el mode d'amplada de banda baixa",
+        "desc": "Reprodueix sempre el canal d'amplada de banda baixa integrat per Frigate en lloc del flux seleccionat. Funciona en qualsevol connexió, però té menor qualitat i sense àudio."
+      }
     },
     "playInBackground": {
       "label": "Reproduir en segon pla",
       "tips": "Habilita aquesta opció per a contiuar la transmissió tot i que el reproductor estigui ocult."
     },
     "debug": {
-      "picker": "Selecció de stream no disponible en mode debug. La vista debug sempre fa servir el stream assignat pel rol de detecció."
+      "picker": "Selecció de stream no disponible en mode debug. La vista debug sempre fa servir el stream assignat pel rol de detecció.",
+      "technology": "La selecció de la tecnologia de flux no està disponible en el mode de depuració."
+    },
+    "mode": "Tecnologia Streaming",
+    "technology": {
+      "description": "Tria la teva tecnologia de streaming preferida. Frigate encara pot tornar al mode d'amplada de banda baixa en la reproducció o els errors de xarxa.",
+      "tips": {
+        "mse": "Predeterminada recomanada, compatibilitat àmplia i reproducció suau",
+        "webrtc": "Necessita configuració addicional i no està suportat en tots els dispositius"
+      },
+      "unavailable": {
+        "browser": "El teu navegador no suporta WebRTC.",
+        "not-configured": "WebRTC no està configurat. Estableix candidats go2rtc webrtc o ice_servers.",
+        "unreachable": "WebRTC no s'ha pogut connectar. Comproveu que el port 8555 és accessible i qualsevol servidor STUN/TURN és correcte.",
+        "video-codec": "El còdec de vídeo d'aquest flux no és compatible amb WebRTC en aquest navegador.",
+        "audio-codec": "El còdec d'àudio d'aquest flux no és compatible amb WebRTC. Transcodifica a opus o G.711 amb go2rtc.",
+        "checking": "S'està comprovant la disponibilitat de WebRTC…"
+      },
+      "name": {
+        "mse": "MSE",
+        "webrtc": "WebRTC",
+        "jsmpeg": "JSMpeg"
+      }
     }
   },
   "streamingSettings": "Paràmetres de transmissió",

web/public/locales/ca/views/settings.json

@@ -88,7 +88,7 @@
     "uiSettings": "Paràmetres de la IU",
     "profiles": "Perfils",
     "systemGo2rtcStreams": "go2rtc streams",
-    "systemDetectorsAndModel": "Detectors i model"
+    "systemDetectorsAndModel": "Models de detecció"
   },
   "dialog": {
     "unsavedChanges": {
@@ -150,11 +150,50 @@
     "toast": {
       "success": {
         "clearStoredLayout": "Disposició emmagatzemada esborrada per {{cameraName}}",
-        "clearStreamingSettings": "S'han suprimit els paràmetres de la transmissió per tots els grups de càmeres."
+        "clearStreamingSettings": "S'han suprimit els paràmetres de la transmissió per tots els grups de càmeres.",
+        "exportUiSettings": "Configuració exportada",
+        "importUiSettings": "Configuració importada"
       },
       "error": {
         "clearStoredLayoutFailed": "Error en suprimir la disposició desada: {{errorMessage}}",
-        "clearStreamingSettingsFailed": "Error en esborrar els paràmetres de la transmissió: {{errorMessage}}"
+        "clearStreamingSettingsFailed": "Error en esborrar els paràmetres de la transmissió: {{errorMessage}}",
+        "exportUiSettingsFailed": "No s'han pogut exportar els paràmetres",
+        "importNothingToApply": "No s'ha pogut importar la configuració: el fitxer no conté cap configuració",
+        "importInvalidJson": "No s'ha pogut importar la configuració: el fitxer no és un JSON vàlid",
+        "importWrongType": "Ha fallat la importació de la configuració: el fitxer no és una exportació de la configuració de Frigate",
+        "importUnsupportedVersion": "No s'ha pogut importar la configuració: el fitxer requereix una versió més nova de Frigate",
+        "importInvalidSchema": "No s'ha pogut importar la configuració: el fitxer està mal format",
+        "importUiSettingsFailed": "No s'han pogut importar els paràmetres"
+      }
+    },
+    "backupRestore": {
+      "title": "Còpia de seguretat i restauració",
+      "transfer": {
+        "label": "Fitxer de configuració del dispositiu",
+        "desc": "Les disposicions del grup de càmeres, la configuració de la transmissió i les preferències de la interfície d'usuari s'emmagatzemen al vostre navegador. Exporta-les a un fitxer per fer-ne una còpia de seguretat o per moure-les a un altre dispositiu.",
+        "export": "Exporta la configuració",
+        "import": "Importa els paràmetres"
+      },
+      "importDialog": {
+        "title": "Importa els paràmetres",
+        "desc": "Trieu què aplicar d'aquest fitxer. Frigate es tornarà a carregar quan acabi la importació.",
+        "exportedFrom": "S'ha exportat {{date}} de la versió de configuració de Frigate {{version}}",
+        "layouts_one": "Distribucions de grup de càmeres (grup{{count}})",
+        "layouts_many": "Distribucions de grup de càmera ({{count}} grups)",
+        "layouts_other": "Distribucions de grup de càmera ({{count}} grups)",
+        "streaming_one": "Configuració de l'Streaming (càmera {{count}})",
+        "streaming_many": "Configuració de l'Streaming (càmeres {{count}})",
+        "streaming_other": "Configuració de l'Streaming (càmeres {{count}})",
+        "preferences_one": "Preferències de la interfície d'usuari (ajust {{count}})",
+        "preferences_many": "Preferències de la interfície d'usuari ( {{count}} paràmetres)",
+        "preferences_other": "Preferències de la interfície d'usuari ( {{count}} paràmetres)",
+        "unknownGroups_one": "El grup de càmeres {{groups}} no és en aquest servidor. La seva configuració es desarà però no s'utilitzarà.",
+        "unknownGroups_many": "Els grups de càmeres {{groups}} no són en aquest servidor. La seva configuració es desarà però no s'utilitzarà.",
+        "unknownGroups_other": "Els grups de càmeres {{groups}} no són en aquest servidor. La seva configuració es desarà però no s'utilitzarà.",
+        "unknownCameras_one": "La càmera {{cameras}} no és en aquest servidor. S'ignorarà la seva configuració de transmissió.",
+        "unknownCameras_many": "Les càmeres {{cameras}} no són en aquest servidor. S'ignoraran els seus paràmetres de transmissió.",
+        "unknownCameras_other": "Les càmeres {{cameras}} no són en aquest servidor. S'ignoraran els seus paràmetres de transmissió.",
+        "confirm": "Importa"
       }
     }
   },
@@ -1024,7 +1063,7 @@
       }
     },
     "step3": {
-      "description": "Configura els rols de flux i afegeix fluxos addicionals per a la càmera.",
+      "description": "Configura característiques, rols de flux i afegeix fluxos addicionals per a la càmera.",
       "streamTitle": "Flux {{number}}",
       "roles": "Rols",
       "streamsTitle": "Fluxos de la càmera",
@@ -1058,11 +1097,26 @@
         "title": "Roles de flux",
         "detect": "Canal principal per a la detecció d'objectes.",
         "record": "Desa els segments del canal de vídeo basats en la configuració.",
-        "audio": "Canal per a la detecció basada en àudio."
+        "audio": "Canal per a la detecció basada en àudio.",
+        "record_sub": "Enregistraments de menor qualitat per a la reproducció adaptativa i la retenció estesa (típicament el substream de la càmera)."
       },
       "featuresPopover": {
         "title": "Característiques del flux",
         "description": "Utilitzeu el restreaming go2rtc per reduir les connexions a la càmera."
+      },
+      "appleCompatibility": {
+        "title": "Millora la reproducció en dispositius Apple",
+        "description": "Activeu això si mireu enregistraments al Safari o a un iPhone, iPad o Mac."
+      },
+      "ptz": {
+        "title": "Habilita els controls PTZ",
+        "detectedNote": "S'ha detectat suport PTZ a través de l'ONVIF. Si es tracta d'una càmera PTZ, habilitar aquesta opció us permetrà controlar les funcions pan/tilt/zoom de la càmera des de la interfície d'usuari.",
+        "connectionDetails": "Detalls de la connexió ONVIF",
+        "host": "Amfitrió ONVIF",
+        "port": "Port ONVIF",
+        "username": "Nom d'usuari ONVIF",
+        "password": "Contrasenya ONVIF",
+        "hostRequiredWarning": "Es requereix un amfitrió i un port ONVIF quan s'habiliten els controls PTZ."
       }
     },
     "step4": {
@@ -1393,7 +1447,9 @@
         "orphansDeleted": "Orfes eliminats",
         "aborted": "Avortat. La supressió superaria el llindar de seguretat.",
         "error": "Error",
-        "totals": "Totals"
+        "totals": "Totals",
+        "spaceToReclaim": "Espai a reclamar",
+        "spaceReclaimed": "Espai reclamat"
       },
       "event_snapshots": "Instantànies de l'objecte rastrejat",
       "event_thumbnails": "Miniatures d'objecte rastrejat",
@@ -1489,7 +1545,8 @@
         "preset-record-mjpeg": "Registre - Càmeres MJPEG",
         "preset-record-jpeg": "Registre - Càmeres JPEG",
         "preset-record-ubiquiti": "Registre - Càmeres Ubiquiti"
-      }
+      },
+      "sameAsRecord": "Igual que els arguments de sortida del registre"
     },
     "cameraInputs": {
       "itemTitle": "Flux {{index}}",
@@ -1570,8 +1627,11 @@
       "options": {
         "detect": "Detecta",
         "record": "Enregistrament",
-        "audio": "Àudio"
-      }
+        "audio": "Àudio",
+        "record_sub": "Registre (Sub Flux)"
+      },
+      "roleInUse": "Ja assignat a un altre flux",
+      "recordSubConflict": "Un flux no pot tenir els rols record andsub"
     },
     "review": {
       "title": "Configuració de la revisió"
@@ -1593,7 +1653,8 @@
       "options": {
         "embeddings": "Vectors",
         "descriptions": "Descripcions",
-        "chat": "Xat"
+        "chat": "Xat",
+        "transcribe": "Transcripció"
       }
     },
     "semanticSearchModel": {
@@ -1618,7 +1679,12 @@
     },
     "knownPlates": {
       "namePlaceholder": "per exemple. Cotxe de la parella",
-      "platePlaceholder": "Matricula o regex"
+      "platePlaceholder": "Matricula o regex",
+      "assignedTo": "Assignat a {{name}}",
+      "detected": "Plaques detectades",
+      "noneDetected": "Encara no s'ha detectat cap matrícula",
+      "search": "Cerca o introdueix una matrícula",
+      "useCustom": "Utilitza \"{{value}}\""
     },
     "semanticSearchModelSize": {
       "notApplicable": "No aplicable als proveïdors de GenAI"
@@ -1643,7 +1709,26 @@
     },
     "defaultRole": {
       "admin": "Administrar",
-      "viewer": "Visor"
+      "viewer": "Visor",
+      "none": "Cap (denega l'accés)"
+    },
+    "birdseyeModes": {
+      "summary": "{{count}} seleccionats",
+      "options": {
+        "continuous": "Continu",
+        "motion": "Moviment",
+        "all_objects": "Tots els objectes",
+        "alerts": "Alertes",
+        "detections": "Deteccions"
+      }
+    },
+    "audioTranscriptionModel": {
+      "placeholder": "Selecciona el model…",
+      "builtIn": "Models integrats",
+      "genaiProviders": "Proveïdors de GenAI"
+    },
+    "audioTranscriptionModelSize": {
+      "notApplicable": "No aplicable als proveïdors de GenAI"
     }
   },
   "globalConfig": {
@@ -1715,7 +1800,10 @@
     "overriddenBaseConfigHeading_one": "El perfil {{profile}} substitueix el camp {{count}} de la configuració base:",
     "overriddenBaseConfigHeading_many": "El perfil {{profile}} substitueix {{count}} camps de la configuració base:",
     "overriddenBaseConfigHeading_other": "El perfil {{profile}} substitueix {{count}} camps de la configuració base:",
-    "overriddenBaseConfigNoDeltas": "El perfil {{profile}} substitueix aquesta secció, però no hi ha valors de camp diferents de la configuració base."
+    "overriddenBaseConfigNoDeltas": "El perfil {{profile}} substitueix aquesta secció, però no hi ha valors de camp diferents de la configuració base.",
+    "overriddenLive": "Sobreescrit (Viure)",
+    "overriddenLiveTooltip": "Aquesta càmera s'està executant amb un valor diferent del desat a la configuració, que es mostra aquí. La vista en viu, MQTT, o un perfil actiu pot canviar-la mentre s'està executant Frigate.",
+    "overriddenLiveValue": "Valor en execució: {{value}}"
   },
   "profiles": {
     "title": "Perfils",
@@ -1833,13 +1921,18 @@
       "recordDisabled": "L'enregistrament està desactivat, els elements de revisió no es generaran.",
       "detectDisabled": "La detecció d'objectes està desactivada. Els elements de revisió requereixen objectes detectats per categoritzar alertes i deteccions.",
       "allNonAlertDetections": "Totes les activitats que no siguin alertes s'inclouran com a deteccions.",
-      "genaiImageSourceRecordingsRecordDisabled": "La font d'imatges està configurada com a 'enregistraments', però l'enregistrament està desactivat. Frigate tornarà a la vista prèvia de les imatges."
+      "genaiImageSourceRecordingsRecordDisabled": "La font d'imatges està configurada com a 'enregistraments', però l'enregistrament està desactivat. Frigate tornarà a la vista prèvia de les imatges.",
+      "recordRuntimeDisabled": "L'enregistrament està activat a la configuració, però actualment està desactivat per a aquesta càmera, de manera que els elements de revisió no es generaran. Torneu-lo a activar des de la vista en viu de la càmera, o comproveu si un perfil actiu l'està desactivant.",
+      "detectRuntimeDisabled": "La detecció d'objectes està activada a la configuració, però actualment està desactivada per a aquesta càmera. Els elements de revisió requereixen objectes detectats per categoritzar alertes i deteccions. Torneu-lo a activar des de la vista en viu de la càmera, o comproveu si un perfil actiu l'està desactivant.",
+      "genaiImageSourceRecordingsRecordRuntimeDisabled": "La font d'imatges està configurada com a 'enregistraments', però l'enregistrament està desactivat per a aquesta càmera tot i que la configuració l'habilita. Frigate tornarà a la vista prèvia de les imatges."
     },
     "audio": {
       "noAudioRole": "Cap flux té definit el rol d'àudio. Heu d'habilitar el rol d'àudio per a la detecció d'àudio perquè funcioni."
     },
     "audioTranscription": {
-      "audioDetectionDisabled": "La detecció d'àudio no està activada per a aquesta càmera. La transcripció d'àudio requereix que la detecció d'àudio estigui activa."
+      "audioDetectionDisabled": "La detecció d'àudio no està activada per a aquesta càmera. La transcripció d'àudio requereix que la detecció d'àudio estigui activa.",
+      "genaiProviderSelected": "La detecció d'àudio està activada a la configuració, però actualment està desactivada per a aquesta càmera, de manera que la transcripció d'àudio no s'executarà. Torneu-lo a activar des de la vista en viu de la càmera, o comproveu si un perfil actiu l'està desactivant.",
+      "audioDetectionRuntimeDisabled": "La detecció d'àudio està activada a la configuració, però actualment està desactivada per a aquesta càmera, de manera que la transcripció d'àudio no s'executarà. Torneu-lo a activar des de la vista en viu de la càmera, o comproveu si un perfil actiu l'està desactivant."
     },
     "detect": {
       "fpsGreaterThanFive": "No es recomana establir el detect FPS superior a 5. Els valors més alts poden causar problemes de rendiment i no proporcionaran cap benefici.",
@@ -1849,7 +1942,9 @@
       "maxFramesSet": "La configuració dels fotogrames màxims anul·la el comportament predeterminat i desactiva el seguiment d'objectes estacionaris. Hi ha molt poques situacions en què això sigui necessari, utilitzeu-lo amb precaució.",
       "squareResolution": "Una resolució de detecció quadrada és inusual. L'amplada i l'alçada de la detecció han de coincidir amb la relació d'aspecte de la càmera (per exemple, 16:9), no amb les dimensions del model de detecció d'objectes. Una relació d'aspecte no coincident pot estirar la imatge i reduir la precisió de detecció.",
       "resolutionHigh": "Aquesta resolució de detecció és més alta del recomanat i pot causar un ús més elevat dels recursos sense millorar la precisió de detecció. Es recomana una resolució de detecció a o per sota de 1080p per a la majoria de les càmeres.",
-      "globalResolutionMultipleCameras": "S'estableix una resolució de detecció global mentre es configuren diverses càmeres. Tret que totes les càmeres comparteixin la mateixa resolució i relació d'aspecte, l'amplada i l'alçada de la detecció s'haurien de definir per càmera perquè coincideixi amb la relació d'aspecte nativa de cada càmera."
+      "globalResolutionMultipleCameras": "S'estableix una resolució de detecció global mentre es configuren diverses càmeres. Tret que totes les càmeres comparteixin la mateixa resolució i relació d'aspecte, l'amplada i l'alçada de la detecció s'haurien de definir per càmera perquè coincideixi amb la relació d'aspecte nativa de cada càmera.",
+      "sceneWithoutModel": "No s'ha configurat cap model de detecció per a aquesta escena, de manera que aquesta càmera torna al model amb una escena de 'Totes les càmeres'. Afegiu un model per a aquesta escena per donar la càmera pròpia.",
+      "runtimeDisabled": "La detecció d'objectes està activada a la configuració, però actualment està desactivada per a aquesta càmera. Instantànies, articles de revisió i enriquiments com el reconeixement de rostres, el reconeixement de matrícules i la IA Generativa no funcionaran fins que no s'encengui des de la vista en viu de la càmera, o fins que el perfil actiu deixi d'inhabilitar-lo."
     },
     "faceRecognition": {
       "globalDisabled": "L'enriquiment del reconeixement facial s'ha d'habilitar perquè les funcions de reconeixement facial funcionin en aquesta càmera.",
@@ -1862,10 +1957,12 @@
       "modelSizeLarge": "El model 'gran' està optimitzat per a matrícules multilínies. El model 'petit' proporciona un millor rendiment sobre 'gran' i s'ha d'utilitzar tret que la vostra regió utilitzi formats de placa multilínia."
     },
     "record": {
-      "noRecordRole": "Cap flux té el rol de registre definit. L'enregistrament no funcionarà."
+      "noRecordRole": "Cap flux té el rol de registre definit. L'enregistrament no funcionarà.",
+      "noRecordSubRole": "Cap flux té definit el rol recordsubsub. L'enregistrament del substream no funcionarà."
     },
     "snapshots": {
-      "detectDisabled": "La detecció d'objectes està desactivada. Les instantànies es generen a partir d'objectes rastrejats i no es crearan."
+      "detectDisabled": "La detecció d'objectes està desactivada. Les instantànies es generen a partir d'objectes rastrejats i no es crearan.",
+      "detectRuntimeDisabled": "La detecció d'objectes està activada a la configuració, però actualment està desactivada per a aquesta càmera, de manera que no es crearan instantànies. Torneu-lo a activar des de la vista en viu de la càmera, o comproveu si un perfil actiu l'està desactivant."
     },
     "objects": {
       "genaiNoDescriptionsProvider": "Heu de configurar un proveïdor de GenAI amb el rol 'descripcions' per a les descripcions que es generaran."
@@ -1883,6 +1980,10 @@
     "model": {
       "optimizedFor320": "Frigate està optimitzada per a un model 320x320, que és la millor opció per a la majoria de configuracions. Un model 640x640 és més lent i només ajuda en escenaris específics.",
       "inputDimensionsNotDetectResolution": "L'amplada i l'alçada del model d'entrada són les dimensions d'entrada del model de detecció d'objectes, no la resolució de detecció de la càmera. Haurien de coincidir amb les dimensions del model que esteu utilitzant - típicament una mida quadrada com 320x320 o 640x640."
+    },
+    "birdseye": {
+      "objectTrackingDetectDisabled": "Birdseye inclou objectes rastrejats, però la detecció d'objectes està desactivada per a aquesta càmera. La càmera no apareixerà a Birdseye.",
+      "objectTrackingDetectRuntimeDisabled": "La detecció d'objectes està activada a la configuració, però actualment està desactivada per a aquesta càmera. Instantànies, articles de revisió i enriquiments com el reconeixement de rostres, el reconeixement de matrícules i la IA Generativa no funcionaran fins que no s'encengui des de la vista en viu de la càmera, o fins que el perfil actiu deixi d'inhabilitar-lo."
     }
   },
   "modelSize": {
@@ -1913,6 +2014,10 @@
     "imageSource": {
       "recordings": "Gravacions",
       "previews": "Previsualitzacions"
+    },
+    "frameMode": {
+      "frames": "Fotogrames",
+      "annotated_frames": "Fotogrames anotats"
     }
   },
   "logger": {
@@ -1940,5 +2045,46 @@
     "overrideGlobal": "Aquesta secció substitueix la configuració global",
     "overrideProfile": "Aquesta secció està substituïda pel perfil {{profile}}",
     "unsaved": "Aquesta secció té canvis sense desar"
+  },
+  "detectionModels": {
+    "title": "Models de detecció",
+    "description": "Configura els models de detecció d'objectes i el maquinari que s'executa. Les càmeres trien un model segons la seva escena a la configuració de detecció de la càmera.",
+    "addModel": "Afegeix un model",
+    "cameras_one": "{{count}} càmera",
+    "cameras_many": "{{count}} càmeres",
+    "cameras_other": "{{count}} càmeres",
+    "scene": {
+      "label": "Escena",
+      "description": "L'entorn per al qual serveix aquest model. Les càmeres trien un model establint la mateixa escena a la seva configuració de detecció, i el model amb una escena de tots és utilitzat per qualsevol càmera que no en estableixi una."
+    },
+    "scenes": {
+      "all": "Totes les càmeres",
+      "indoor": "Interior",
+      "outdoor": "Exterior",
+      "indoor_thermal": "Tèrmiques interiors",
+      "outdoor_thermal": "Tèrmiques exterior"
+    },
+    "hardware": {
+      "label": "Maquinari",
+      "placeholder": "Selecciona el maquinari",
+      "loading": "S'està cercant el maquinari de detecció...",
+      "none": "Cap maquinari seleccionat",
+      "claimedBy": "utilitzat per {{scene}}",
+      "detectorCount": "Detectors",
+      "countRecommended_one": "Recomanat per a la càmera {{count}}",
+      "countRecommended_many": "Recomanat per {{count}} càmeres",
+      "countRecommended_other": "Recomanat per {{count}} càmeres",
+      "unrecognized": "Aquest model està configurat per al maquinari que no s'ha trobat en aquest sistema: {{devices}}",
+      "description": "El maquinari en el qual aquest model executa la seva detecció.",
+      "detectorCountDescription": "Quants processos de detecció s'han d'executar en aquest maquinari. Més detectors es mantenen al dia amb més càmeres, a costa de la memòria extra del dispositiu.",
+      "unitsDescription": "Cada unitat executa el seu propi procés de detecció. No es pot seleccionar una unitat ja utilitzada per un altre model."
+    },
+    "tabs": {
+      "plus": "Frigate+",
+      "custom": "Model personalitzat"
+    },
+    "plusModel": {
+      "noModelSelected": "Selecciona un model Frigate+"
+    }
   }
 }

web/public/locales/ca/views/system.json

@@ -9,7 +9,8 @@
       "nginx": "Registres de Nginix - Frigate",
       "websocket": "Registres de missatges - Frigate"
     },
-    "enrichments": "Estadístiques complementàries - Frigate"
+    "enrichments": "Estadístiques complementàries - Frigate",
+    "health": "Salut - Frigate"
   },
   "title": "Sistema",
   "metrics": "Mètriques del sistema",
@@ -139,7 +140,11 @@
         "title": "Sense utilitzar",
         "tips": "Aquest valor pot no de forma exacta representar l'espai lliure disponible a Frigate si tens altres fitxers emmagatzemats en la vostra unitat més enllà dels registres de Frigate. Frigate no rastreja l'ús d'emmagatzematge extern als seus registres."
       },
-      "percentageOfTotalUsed": "Percentatge del total"
+      "percentageOfTotalUsed": "Percentatge del total",
+      "subStream": {
+        "information": "Informació de l'emmagatzematge del subflux",
+        "tips": "Les càmeres que enregistren un subflux mostren el seu ús original i de baixa qualitat per separat, utilitzant els mateixos noms que les opcions de qualitat en la reproducció de l'enregistrament. Els enregistraments de baixa qualitat es mantenen en disc fins que el seu període de retenció expira, de manera que aquesta avaria encara pot aparèixer per a una càmera després que s'hagi desactivat l'enregistrament de substream."
+      }
     },
     "overview": "Visió general",
     "shm": {
@@ -197,7 +202,11 @@
         "segmentLength": "Longitud del segment d'enregistrament:",
         "ok": "Fotogrames clau cada ,{{seconds}}s, bons per enregistrar i reproduir.",
         "unknown": "No s'ha pogut determinar l'espaiat dels fotogrames clau.",
-        "recordDisabled": "L'enregistrament està desactivat per a aquesta càmera."
+        "recordDisabled": "L'enregistrament està desactivat per a aquesta càmera.",
+        "warningFixed": "Els fotogrames clau estan uniformement espaiats però dispersos (cada ).{{seconds}}s). L'enregistrament encara funciona, però la reproducció en directe i la recerca comencen més lentament. Estableix l'interval I-frame (keyframe) de la càmera perquè coincideixi amb la seva velocitat de fotogrames.",
+        "warningVariable": "L'espaiat dels fotogrames clau és inconsistent ({{minSeconds}}s a {{maxSeconds}}s), el qual normalment significa que un còdec intel·ligent (H.264+/H.265+) està habilitat. Això no és recomanable.",
+        "errorFixed": "Els fotogrames clau cada the{{seconds}}s és més llarg que la longitud del segment d'enregistrament ({{segmentTime}}s), de manera que alguns segments no tenen fotogrames clau i no es reproduiran. Escurça l'interval I-frame (keyframe) de la càmera perquè coincideixi amb la seva velocitat de fotogrames.",
+        "errorVariable": "Els buits dels fotogrames clau arriben a the{{seconds}}s, més llargs que la longitud del segment d'enregistrament ({{segmentTime}}s). Alguns segments no tindran un fotograma clau, el qual trenca la reproducció. Desactiva el còdec intel·ligent/+ a la càmera o escurça el seu interval de fotogrames clau."
       }
     },
     "title": "Càmeres",
@@ -234,7 +243,11 @@
     "detectHighCpuUsage": "{{camera}} te un ús elevat de CPU per la detecció ({{detectAvg}}%)",
     "detectIsVerySlow": "{{detect}} és molt lent ({{speed}} ms)",
     "detectIsSlow": "{{detect}} és lent ({{speed}} ms)",
-    "debugReplayActive": "La sessió de repetició de depuració està activa"
+    "debugReplayActive": "La sessió de repetició de depuració està activa",
+    "cameraSkippedDetections": "{{camera}} està ometent la detecció en {{pct}}% de fotogrames",
+    "systemNotices_one": "{{count}} avís del sistema",
+    "systemNotices_other": "{{count}} avisos del sistema",
+    "retentionUnmet": "Els enregistraments s'estan suprimint abans que acabi el seu període de retenció a l'espai lliure"
   },
   "enrichments": {
     "title": "Anàlisi avançada",
@@ -262,5 +275,93 @@
     },
     "infPerSecond": "Inferències per segon",
     "averageInf": "Temps mitjà d'inferència"
+  },
+  "health": {
+    "title": "Salut",
+    "notices": {
+      "title": "Avís",
+      "empty": "La instal·lació de Frigate és saludable",
+      "dismiss": "Descarta",
+      "openSettings": "Obre la configuració",
+      "openLink": "Obre l'enllaç",
+      "noMatches": "No hi ha avisos que coincideixin amb el filtre",
+      "dismissedTitle": "Descartat",
+      "noneDismissed": "No hi ha avisos descartats",
+      "clearDismissed": "Neteja descartada",
+      "clearDismissedTitle": "Voleu netejar els avisos descartats?",
+      "clearDismissedDesc": "S'elimina cada avís acomiadat. Les comprovacions de configuració i flux es mostren de nou immediatament, i altres avisos tornen la propera vegada que succeeixin.",
+      "firstSeen_one": "{{time}} vist per primera vegada",
+      "firstSeen_other": "{{time}} · {{count}} vegades",
+      "dismissedAt": "S'ha suprimit {{time}}",
+      "filter": {
+        "showDismissed": "Mostra descartat",
+        "severity": "Severitat",
+        "warning": "Avís",
+        "error": "Error",
+        "info": "Info"
+      },
+      "kinds": {
+        "detector_stuck": "S'ha reiniciat el detector {{detector}} després que deixés de respondre",
+        "model_download_failed": "Ha fallat la baixada de {{file}} per {{model}}: {{error}}",
+        "skipped_detections": "La detecció no ha pogut mantenir el {{pct}}% de fotogrames durant més d'un minut",
+        "shm_too_low": "L'assignació /dev/shm ({{total}} MB) s'hauria d'augmentar a com a mínim {{min}} MB",
+        "failed_login_one": "Error en l'intent d'inici de sessió de {{user}}",
+        "failed_login_other": "Intents d'inici de sessió fallits per {{user}}",
+        "update_available": "Frigate {{version}} està disponible"
+      },
+      "streamPrefix": "Flux {{index}}: {{message}}",
+      "streamPrefixRestream": "Flux {{index}} (via go2rtc): {{message}}",
+      "streamProbeFailed": "No s'ha pogut provar el flux {{index}}: {{error}}",
+      "cameraProbeFailed": "No s'han pogut explorar els fluxos: {{error}}"
+    },
+    "hardware": {
+      "title": "Maquinari",
+      "objectDetection": "Detecció d'objectes",
+      "hardwareAcceleration": "Acceleració del maquinari",
+      "enrichments": {
+        "title": "Enriquiments",
+        "semantic_search": "Cerca semàntica",
+        "face_recognition": "Reconeixement facial",
+        "lpr": "Reconeixement de matricules",
+        "audio_transcription": "Transcripció d'àudio"
+      },
+      "cameraConnections": "Connexions de la càmera",
+      "allCamerasExcellent": "Totes les càmeres tenen una excel·lent connexió.",
+      "fps": "{{camera}} / {{expected}} fps",
+      "nothingConfigured": "Res configurat",
+      "allCameras": "Totes les càmeres",
+      "probeUnavailable": "La detecció de maquinari no està disponible",
+      "justStarted": "Frigate acaba de començar",
+      "deviceNotFound": "{{devices}} no s'ha trobat en aquest sistema",
+      "detectorNotRunning": "El procés del detector no s'està executant",
+      "inferenceVerySlow": "La inferència és molt lenta ({{speed}} ms)",
+      "inferenceSlow": "La inferència és lenta ({{speed}} ms)",
+      "customArgs": "Arguments personalitzats del ffmpeg",
+      "customArgsNotVerified": "Arguments personalitzats del ffmpeg, no verificats",
+      "hwaccelNotConfigured": "No s'utilitza cap descodificació de maquinari, però {{family}} està disponible. Seleccioneu-lo a la configuració del ffmpeg",
+      "hwaccelHardwareMissing": "{{family}} està configurat, però la sonda de maquinari no ho ha informat",
+      "unrecognizedDevice": "Sobreescriu el dispositiu no reconegut, no verificat",
+      "decoderUsage": "descodificador {{usage}}",
+      "remoteProvider": "Proveïdor remot",
+      "acceleratorMissing": "Configurat per {{device}} però no s'ha trobat cap accelerador compatible",
+      "modelNotRunYet": "El model encara no s'ha executat, el dispositiu es verifica després del primer ús",
+      "fellBackToCpu": "Configurat per {{device}} però executant-se a la CPU",
+      "cpuDespiteAccelerator": "S'està executant a la CPU encara que hi ha disponible un accelerador",
+      "probed": "Profund",
+      "probing": "S'està sondejant...",
+      "recheck": "Torna a comprovar el maquinari",
+      "cameraStreams": "Fluxos de la càmera",
+      "streamsNotChecked": "Encara no s'ha comprovat",
+      "streamsChecking": "S'està comprovant {{done}} de {{total}}",
+      "streamsClean": "No hi ha problemes de flux",
+      "checked": "Comprovat",
+      "runAgain": "Torna a executar",
+      "runStreamChecks": "Executa les comprovacions de flux",
+      "streamsChecked_one": "S'ha comprovat la càmera {{count}}",
+      "streamsChecked_other": "{{count}} càmeres comprovades",
+      "streamsFlagged_one": "{{count}} amb problemes, llistat a Avisos",
+      "streamsFlagged_other": "{{count}} amb problemes, llistat a Avisos",
+      "inferenceMs": "{{speed}} ms"
+    }
   }
 }

web/public/locales/cs/components/player.json

@@ -20,14 +20,6 @@
       "title": "Šířka pásma:",
       "short": "Šířka pásma"
     },
-    "latency": {
-      "title": "Latence:",
-      "value": "{{seconds}} sekund",
-      "short": {
-        "title": "Latence",
-        "value": "{{seconds}} sek"
-      }
-    },
     "totalFrames": "Celkový počet snímků:",
     "droppedFrames": {
       "title": "Ztracené snímky:",

web/public/locales/da/audio.json

@@ -343,5 +343,8 @@
   "engine": "Motor",
   "light_engine": "Lille motor",
   "swing_music": "Swing musik",
-  "bluegrass": "Bluegrass"
+  "bluegrass": "Bluegrass",
+  "folk_music": "Folkemusik",
+  "echo": "Ekko",
+  "pulse": "Puls"
 }

web/public/locales/da/components/player.json

@@ -20,14 +20,6 @@
       "title": "Båndbredde:",
       "short": "Båndbredde"
     },
-    "latency": {
-      "title": "Latenstid:",
-      "value": "{{seconds}} sekunder",
-      "short": {
-        "title": "Latenstid",
-        "value": "{{seconds}} sek"
-      }
-    },
     "droppedFrames": {
       "short": {
         "title": "Tabt",

web/public/locales/de/components/player.json

@@ -20,14 +20,6 @@
       "title": "Bandbreite:",
       "short": "Bandbreite"
     },
-    "latency": {
-      "title": "Latenz:",
-      "value": "{{seconds}} Sekunden",
-      "short": {
-        "title": "Latenz",
-        "value": "{{seconds}} s"
-      }
-    },
     "droppedFrames": {
       "short": {
         "title": "Ausgelassen",

web/public/locales/el/components/player.json

@@ -20,14 +20,6 @@
       "title": "Ταχύτητα:",
       "short": "Ταχύτητα"
     },
-    "latency": {
-      "title": "Καθυστέρηση:",
-      "value": "{{seconds}} δευτερόλεπτα",
-      "short": {
-        "title": "Καθυστέρηση",
-        "value": "{{seconds}} δευτερόλεπτα"
-      }
-    },
     "totalFrames": "Συνολικός αριθμός Καρέ:",
     "droppedFrames": {
       "title": "Απορριφθέντα καρέ:",

web/public/locales/en/config/cameras.json

@@ -657,6 +657,10 @@
         "label": "Preferred language",
         "description": "Preferred language to request from the GenAI provider for generated responses."
       },
+      "frame_mode": {
+        "label": "Frame mode",
+        "description": "How frames are presented to the model. 'frames' sends the prompt followed by the frames, which suits models that track a sequence well on their own. 'annotated_frames' labels each frame and interleaves notes derived from object tracking, which helps models that lose track of activity that repeats or reverses."
+      },
       "response_style": {
         "label": "Response style",
         "description": "Writing style preset for generated review descriptions. Presets adjust the tone and level of detail of the user-facing title, summary, and scene description; 'default' leaves the built-in prompt unchanged."