openWakeWord (Recommended)
Why This Option
- Fully open source
- No cloud dependency
- No expiration on custom wake words
- Used by Home Assistant community
- Lightweight enough for Raspberry Pi
Installation
pip install openwakeword
GitHub
https://github.com/dscripka/openWakeWord
Training Custom Wake Word
- Use Google Colab notebook (provided in repo)
- Takes approximately 1 hour
- Train "Hey LARS" specifically
- Output: .tflite model file
Performance
- Runs 15-20 models on single RPi 3 core
- Low CPU usage when idle
- Sub-second detection latency
Basic Usage
from openwakeword.model import Model
# Load model
model = Model()
# From audio file
model.predict_clip("path/to/audio.wav")
# From microphone stream
# See repo examples for streaming implementation
Integration with Home Assistant
- Place .tflite file in /share/openwakeword
- Configure in voice assistant settings
- Works with Wyoming protocol
Next Steps
- Train "Hey LARS" wake word model
- Test detection accuracy
- Integrate with LiveKit or custom audio pipeline