Silero VAD

Silero VAD (snakers4/silero-vad, active since 2020, now at v5) is the open-source voice activity detector that quietly became the industry default — it sits inside faster-whisper, LiveKit, sherpa-onnx and countless voice pipelines. The recipe is deliberately boring in the best way: a small LSTM-based network (~309K params in v5, ~2MB JIT/ONNX) that takes 16kHz or 8kHz audio in 256/512-sample chunks and emits a speech probability per chunk, with hysteresis-style post-processing to turn probabilities into clean speech timestamps. The provider's own benchmarks show it comfortably ahead of the classic WebRTC VAD on noisy speech, and an independent 2026 comparison table reports 87.7% true-positive rate at 5% false-positive rate (WebRTC VAD: 50.0%). There are no strings attached — MIT license, no telemetry, no keys, runs anywhere PyTorch or ONNX runs. The main knobs are the decision threshold and minimum speech/silence durations; push the threshold down if you're clipping first words, up if background chatter leaks through.

• <1ms per 30ms+ audio chunk on a single CPU thread (provider-reported)\n- 87.7% TPR at 5% FPR on LibriSpeech+DEMAND-style benchmark (third-party 2026 comparison; WebRTC VAD 50.0%, Cobra 98.9% on the same table)\n- Provider wiki publishes quality metrics vs other VAD solutions