Silero VAD is a pre-trained voice activity detection model: given an audio stream, it returns the timestamps where someone is speaking. Voice agents, telephony bots, and transcription pipelines use it to decide when speech starts and stops and to drop silence before audio reaches a speech-to-text model.
The model ships as a small PyTorch JIT model (about 2 MB) and as an ONNX model, and runs wherever those runtimes run. A Python package exposes helpers such as get_speech_timestamps, and the model only needs PyTorch when the caller already loads audio itself. It was trained on corpora covering more than 6,000 languages and audio from varied domains, noise levels, and quality.
Processing one audio chunk of 30 ms or more takes under 1 ms on a single CPU thread, and batching, GPUs, or ONNX can raise throughput further. It supports 8 kHz and 16 kHz sampling rates and needs no telemetry, API keys, or registration. Typical uses include edge and mobile devices, data cleaning, call-center automation, voice bots, and voice interfaces.
Features
- Speech timestamps: returns speech segments in samples or seconds for an audio tensor or file
- Small model: the JIT model is about 2 MB and processes a 30 ms chunk in under 1 ms on one CPU thread
- Multilingual training: trained on data spanning more than 6,000 languages and varied noise conditions
- Sampling rates: 8 kHz and 16 kHz input
- Runtimes: PyTorch, ONNX Runtime on CPU or GPU, ExecuTorch, OpenVINO conversion, and a tinygrad model
- Language examples: usage examples in C++, Rust, Go, Java, C#, and in the browser through ONNX Runtime Web
- Long recordings: optional offline ONNX sequence inference for long audio files
- No account needed: no telemetry, keys, registration, or built-in expiration
