TEN VAD is a voice activity detector designed for real-time conversational AI, offering low latency and high performance. It targets developers building voice-enabled applications, providing superior accuracy and efficiency compared to common alternatives like WebRTC VAD and Silero VAD.

How It Works

TEN VAD employs a proprietary architecture optimized for temporal efficiency, enabling rapid speech-to-non-speech transition detection. This approach minimizes end-to-end latency in conversational AI systems and effectively handles short silences between speech segments, a common failure point for other VADs.

Quick Start & Requirements

• Installation: git clone https://github.com/TEN-framework/ten-vad.git
• Python Usage: pip install -r requirements.txt (for examples/plotting), pip install -U --force-reinstall -v git+https://github.com/TEN-framework/ten-vad.git (for direct use).
• Dependencies: Python (3.8.19/3.10.14 verified), numpy, scipy, scikit-learn, matplotlib, torchaudio. ONNX usage requires onnxruntime >= 1.17.1. C/C++ usage requires Clang/Visual Studio/Xcode and CMake.
• Platforms: Linux, Windows, macOS, Android, iOS, Web (WASM/JS).
• Resources: Setup time varies by platform; core library size is lightweight (e.g., 306KB on Linux x64).
• Demo: Hugging Face Space: https://github.com/user-attachments/assets/725a8318-d679-4b17-b9e4-e3dce999b298

Highlighted Details

• Achieves superior precision-recall compared to WebRTC VAD and Silero VAD on benchmark datasets.
• Demonstrates significantly lower latency in speech-to-non-speech transitions than Silero VAD.
• Offers substantially lower computational complexity and smaller library size than Silero VAD across multiple platforms.
• Provides cross-platform C compatibility and Python, JS (WASM), Android, and iOS bindings.

Maintenance & Community

• Active development with recent updates integrating into k2-fsa/sherpa-onnx and releasing ONNX models.
• Community channels: Discord, X, LinkedIn, Hugging Face.

Licensing & Compatibility

• Licensed under Apache 2.0.
• Includes code derived from LPCNet, which is BSD-2-Clause and BSD-3-Clause licensed (details in NOTICES file).

Limitations & Caveats

• Requires resampling to 16kHz for audio inputs at other sampling rates.
• The default threshold of 0.5 may require tuning for specific applications.
• iOS usage requires manual framework embedding and device signature configuration in Xcode.