## Summary
- Introduces configurable wake word and stop word sensitivity
- Adds UI-controllable thresholds for wake word detection
- Refactors wake word loading and handling logic
- Improves audio preprocessing and runtime behavior
## Key Features
- Per-wake-word sensitivity configuration (Wake Word 1 & 2)
- Configurable stop word sensitivity
- Fine-grained numeric control (0.0 – 1.0, step 0.001)
- Sensitivity values exposed as Number Entities for Home Assistant
- Persistent storage of user-defined sensitivity values
## Wake Word Handling
- Support for model-specific default thresholds (`probability_cutoff`)
- Automatic fallback to defaults if no user value is set
- Separate thresholds for:
- Wake Word 1
- Wake Word 2
- Stop Word
- Unified handling for MicroWakeWord and OpenWakeWord models
## UI / API Integration
- Adds new Home Assistant entities:
- `WakeWord1SensitivityNumberEntity`
- `WakeWord2SensitivityNumberEntity`
- `StopWordSensitivityNumberEntity`
- Entities support:
- Live updates
- Persistence via preferences
- Synchronization with runtime state
- Pushes updated values to Home Assistant on startup
## Refactoring
- Extracted wake word logic into `wake_word.py`
- Model discovery
- Model loading
- Stop word handling
- Simplified `__main__.py` by delegating wake word responsibilities
- Improved separation between configuration and runtime state
## Additional Changes
- Improved debug logging across initialization and runtime
## Impact
- Enables fine-tuning of wake word detection accuracy
- Reduces false positives and missed triggers
- Improves usability via UI-based configuration
- Provides better flexibility across different environments and microphones
## WebRTC Auto Gain & Noise Suppression
This PR adds support for WebRTC Auto Gain and Noise Suppression as previously implemented in the Wyoming Satellite project.
### What's New
Two new entities have been created a number slider for gain and another one for noise suppression. The processing happens inside of the `webrtc.py` file and is implemented in the `__main__.py` file's `process_audio` function. The slider for gain has a limit of 0-31 and noise suppression has a limit of 0-4.
The Auto Gain and Noise Suppression persist through restarts and are saved as preferences. The values are dynamically chnged without needing a restart.
CLI flags called `--mic-auto-gain` and `--mic-noise-suppression` have also been added.
> When both values are 0, WebRTC processing is bypassed entirely with no performance overhead.
Closes#25
### Testing
When saving the audio recorded by LVA there is a noticeable increase. The following two files have been recorded with similar audio being played through them with these results (Gain-31 and NS-4):
| Metric | With AGC+NS | Without |
|---|---|---|
| Max amplitude | 0.329 | 0.054 |
| RMS amplitude | 0.008905 | 0.002862 |
| Volume adjustment needed | 3.034x | 13.479x |
| Rough frequency | 916 Hz | 705 Hz |
Introduce a new `--output-only` command-line flag and corresponding `ENABLE_OUTPUT_ONLY` environment variable. This allows users to run the assistant in output-only mode, where it only supports announcements. The change includes updates to configuration examples, documentation, and the Docker entrypoint script.
Output only satellite for nodes with only speakers and no microphone.
Supports announcements and playing media only.
---------
Co-authored-by: Florian Asche <github@florian-asche.de>
Breaking Change: From now on the cookie file will be created in the configured directory/file. There will be no alternative directory where we try to create and touch the file. You need to make sure that the configured directory is writeable from the application.
Wake-up after stopping ringing timer using wake words such as "Okay Nabu", "Hey Jarvis" rather than "Stop" to get another command instead of the need to repeat the wake word.
Auto stop ringing timer if the determined time (default 15 mins) has passed without manually stopping. This max time can be modified using the command `--timer-max-ring-seconds`.
tts_player.stop() already invokes its done_callback which is
_tts_finished(), so the explicit self._tts_finished() call immediately
after was a duplicate. This caused double 'Unducking music' and double
'TTS response finished' log entries on every stop-word-triggered stop.
Also move _pipeline_active = False inside the 'if not self._tts_played'
branch in VOICE_ASSISTANT_RUN_END, so it is not cleared while TTS is
still playing.
Co-authored-by: Omar Amin <65174210+omaramin-2000@users.noreply.github.com>
Omar Amin