0 Fork 0

Files

T

Omar Amin 3ea32895ee Wake word sensitivity (#207 )

## 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

2026-04-21 02:04:11 +02:00

3.8 KiB

Raw Blame History

Audio Options: Gain, Noise Suppression, and Wake Word Sensitivity

What do Gain and Noise Suppression do

Gain

When you change gain, LVA(using the webrtc library) automatically adjusts the microphone input volume to keep it at a consistent level. If you're speaking quietly it boosts the signal, if you're speaking loudly it reduces it. Useful in the following ways:-

Microphones that are too quiet by default
Environments where you move around relative to the mic
Keeping wake word detection consistent regardless of speaking volume

Noise Suppression

When you change noise suppression, LVA(using the webrtc library) filters out constant background noise from the audio signal. It works by learning what "silence" sounds like in your environment and subtracting that from the audio. Useful in the following ways:

Noisy environments
Improving STT accuracy by sending cleaner audio to Home Assistant
Reducing false wake word triggers from background noise

Using Gain and Noise Suppression with LVA

LVA implements gain and Noise Suppression in two ways:

CLI Argument/ENV file
Home Assistant Entity

CLI Argument/ENV file

CLI Argument

Using docker-entrypoint.sh the following flags can be used to set gain and noise suppression:

--mic-auto-gain(ranges from 0-31)
--mic-noise-suppression(ranges from 0-4)

ENV file

You can add/edit these variables in the .env file to set gain and noise suppression:

MIC_AUTO_GAIN="0"
MIC_NOISE_SUPPRESSION="0"

Home Assistant Entity

LVA exposes two slider entities to change these gain and noise suppression from which you can edit the gain and noise suppression at runtime.

💡 Note: Setting the flag and ENV values to 0 turns them off and are not used.

💡 Note: Keep in mind that when the flags are set they will overwrite the previous value in the preferences file and also they will be overridden if the value is changed in Home Assistant(Also applies to the ENV file but the ENV file will always overwrite the last value on startup).

Wake Word Sensitivity

LVA exposes three numeric controls in the Home Assistant device page for fine-grained sensitivity tuning. These let you dial in the exact probability threshold that best matches your microphone quality, room acoustics, and false-activation tolerance.

Entity	Description	Default
Wake Word 1 Sensitivity	Probability cutoff for the primary wake word	From model manifest
Wake Word 2 Sensitivity	Probability cutoff for the secondary wake word (if active)	From model manifest
Stop Word Sensitivity	Probability cutoff for the stop word	From model manifest

Values range from 0.0 to 1.0:

Higher value (e.g. 0.95) → more selective, fewer false activations, may miss quieter or accented speech
Lower value (e.g. 0.50) → more responsive, but more likely to trigger on similar-sounding words The defaults are read directly from each model's .json manifest file (the probability_cutoff field), so bundled models like okay_nabu (0.85), hey_jarvis (0.97), and custom downloaded models all start at their author-tested baseline. Changes made in the Home Assistant UI are persisted in preferences.json and survive restarts.

Tuning guide:

If the wake word rarely activates (misses your voice): lower the value slightly, e.g. by 0.05 steps
If the wake word activates too often (false triggers from TV, music, or similar words): raise the value slightly
Start from the model's default and make small adjustments — a change of 0.05–0.10 is usually enough to notice a difference
Far-field microphones with noise cancellation (e.g. ReSpeaker, Satellite1) generally work well at the default values; basic USB microphones may need a lower threshold

3.8 KiB Raw Blame History Unescape Escape