TEN-framework/ten-vad

What it solves

TEN VAD is a real-time voice activity detection (VAD) system designed for enterprise-grade conversational AI. It solves the problem of accurately identifying when a user is speaking versus when there is silence or background noise, specifically aiming to reduce the end-to-end response latency and improve turn-detection precision in human-agent interactions.

How it works

TEN VAD provides frame-level speech activity detection on 16kHz audio input. It uses a lightweight architecture that can be deployed as a prebuilt dynamic library or via an ONNX model. The system processes audio frames (with configurable hop sizes like 10ms or 16ms) and outputs a probability and a binary indicator (speech or non-speech) based on a tunable threshold.

Who it’s for

This tool is designed for developers building conversational voice AI agents who need a high-performance, low-latency VAD that can run across various platforms, including servers (Linux, Windows, macOS) and edge devices (Android, iOS, Web via WASM).

Highlights

• High Precision: Outperforms WebRTC VAD and Silero VAD on manually annotated testsets.
• Low Latency: Rapidly detects speech-to-non-speech transitions, reducing the delay in agent responses.
• Lightweight: Significantly lower computational complexity (RTF) and smaller library size (as low as 277KB on Web) compared to Silero VAD.
• Broad Compatibility: Supports Python, C, Java, Go, and JavaScript (WASM) across multiple operating systems and hardware architectures.