Silero VAD

A high-performance, enterprise-grade voice activity detector pre-trained on 100+ languages for real-time edge and server applications.

Silero VAD delivers production-ready voice detection with a footprint under 2 MB. It processes 30ms audio chunks in less than 1ms on a single CPU core (Intel i7 or better). This model handles 100+ languages and varying sampling rates (8kHz or 16kHz) without retraining. Integration is straightforward via PyTorch, ONNX, or C++: developers typically deploy it for speech-to-text pre-processing or real-time telephony systems. It maintains high accuracy even in noisy environments (SNR down to 5dB) by using a deep neural network architecture optimized for low latency.

https://github.com/snakers4/silero-vad