WebRTC VAD

WebRTC VAD is the voice activity detector Google wrote in C for the WebRTC real-time communication stack (the same stack inside Chrome). It's a Gaussian mixture model, not a neural network: each 10/20/30ms frame is split into six frequency bands between 80Hz and 4kHz and scored against adaptive statistical models of speech and noise, with four aggressiveness modes (0=quality through 3=very aggressive) trading missed speech against false alarms. Its enduring appeal is that it needs no model file, uses negligible CPU, and is perfectly deterministic — the same input always gives the same output. Most developers meet it through py-webrtcvad (pip install webrtcvad), libfvad (standalone C extraction, BSD-3), or a WebAssembly build for the browser. The honest limitation: it reacts to energy in the speech bands, so it handles steady noise well but lets through anything loud and voice-like — televisions, colleagues across the room, café chatter, music with vocals. A 2026 third-party comparison put it at 50.0% TPR at 5% FPR versus 87.7% for Silero and 98.9% for Cobra on the same data. For trimming silences from clean recordings before transcription it's still a sound, nearly-free choice; for noisy rooms, the neural options win.

• 50.0% TPR at 5% FPR on LibriSpeech+DEMAND-style benchmark (third-party 2026 comparison table; Silero 87.7%, Cobra 98.9% on the same table)\n- Negligible CPU: RTF ~0.0002 on desktop CPU (third-party benchmark)\n- No peer-reviewed benchmark; figures come from community/provider-built comparisons