Why it matters
Good VAD makes voice apps feel responsive. Bad VAD wastes compute or feels laggy.
Advertisement
The architecture
Signal analysis: energy, spectral features distinguish speech from noise.
Neural VAD: modern ML models robust to noise.
Advertisement
How it works end to end
WebRTC VAD: lightweight built into WebRTC. Works well for clear voice; less good in noise.
Silero VAD: PyTorch model. Excellent noise robustness. Popular open choice.
Sensitivity tuning: threshold between speech and non-speech. Too sensitive: false triggers. Too strict: cut speech starts.
Endpoint detection: know when speaker is done, for turn-taking.