Silero VAD

Silero VAD answers one narrow question, whether a chunk of audio contains speech, and does it well enough to have become the default speech gate across the open source audio ecosystem. The JIT model weighs about 2 MB and classifies a chunk of 30 ms or more in under 1 ms on a single CPU thread, so it runs comfortably inside realtime pipelines, on edge hardware, and at telephony scale. Training drew on corpora spanning more than 6,000 languages, which makes it robust to accents, noise, and domain shifts, and it accepts 8 kHz and 16 kHz sampling rates. Installation is pip install silero-vad, torch.hub, or the ONNX export, with community ports for C++, Rust, Go, Java, C#, and the browser, plus OpenVINO and ExecuTorch paths. Downstream projects such as RealtimeSTT build it in as their voice activity layer, and countless transcription and voice agent stacks trigger their ASR through it. The model is MIT licensed with no telemetry or vendor lock-in, the repository holds around 9,800 stars, and documentation lives in the project wiki.