A 2 MB model that tells you when someone is actually talking

What it does

Silero VAD spots speech in audio streams. Feed it a 30 ms chunk and it returns whether someone is talking. It handles 8 kHz and 16 kHz inputs, works in Python via PyTorch or ONNX, and ships with wrappers for C++, Rust, Go, Java, C#, and the browser.

The interesting bit

The model was trained on corpora spanning over 6000 languages, which is either thorough or slightly unhinged. The JIT-compiled model weighs about two megabytes and processes faster than real-time on a single CPU thread. The README claims ONNX can run 4–5× faster “under certain conditions,” which is admirably hedged.

Key highlights

• Runs fully offline with no registration, keys, telemetry, or built-in expiration (MIT license)
• Single audio chunk processes in under 1 ms on one CPU thread per the project’s performance wiki
• ONNX Runtime path enables deployment on any architecture onnxruntime supports, including edge and mobile
• Community ports cover browser (ONNX Runtime Web), ExecuTorch, and half a dozen systems languages
• Designed for telephony, IoT, call-center bots, and general data cleaning pipelines

Caveats

• Python examples require a modern x86-64 CPU with AVX/AVX2/AVX-512/AMX and 1 GB+ RAM; non-x86 deployments must implement their own I/O and adapt the wrappers
• Audio I/O depends on torchaudio, which needs FFmpeg, sox, or soundfile installed separately

Verdict

Worth a look if you need lightweight, offline speech detection without cloud dependencies. Skip it if you already have a VAD baked into your speech-to-text stack and don’t need the portability.