Silero VAD

Silero VAD

Silero VAD is a production-ready voice activity detection model trained on thousands of hours of diverse speech. At under 2MB and ~1ms latency per chunk, it is the standard go-to VAD for real-time speech pipelines — used inside faster-whisper, LiveKit, and countless production STT deployments.

Key Features

• Tiny footprint: ~1.8MB ONNX model; runs on CPU in real time
• High accuracy: outperforms WebRTC VAD on diverse speech corpora with far fewer false negatives
• Streaming-first: chunk-by-chunk API designed for live microphone input
• Framework flexibility: PyTorch hub, ONNX, and TensorFlow Lite exports
• Language agnostic: trained on 100+ languages without language-specific tuning
• Timestamps: returns precise speech segment start/end times for downstream chunking

Quick Start

pip install silero-vad

from silero_vad import load_silero_vad, read_audio, get_speech_timestamps

model = load_silero_vad()
wav = read_audio('audio.wav')  # 16kHz mono
timestamps = get_speech_timestamps(wav, model, return_seconds=True)
for ts in timestamps:
    print(f"speech: {ts['start']:.2f}s – {ts['end']:.2f}s")

npx ai-supply add silero-vad-voice-activity-detection

Curated mirror of the open-source Silero VAD (MIT). Get it from the source.