TEN VAD is a native C library from the TEN Framework that provides frame-level speech probability scores with a fixed 256-sample hop size (16 ms at 16 kHz).
Provider identifier: ten_vad
Source Location
api/assistant-api/internal/vad/internal/ten_vad/
├── ten_vad.go # Main implementation
├── detector.go # CGO wrapper for libten_vad
├── ten_vad.h # C header
├── lib/
│ └── Linux/x64/libten_vad.so # Shared library
How It Works
- Incoming LINEAR16 bytes are converted to
int16 samples
- Samples are processed in fixed 256-sample frames (16 ms each)
- Each frame produces a speech probability score
- Speech onset/offset is tracked with the same hysteresis logic as Silero — speech ends only when probability drops below
threshold - 0.15
- Same packet emission pattern:
InterruptionPacket on onset, VadSpeechActivityPacket heartbeats during speech
Parameters
| Option Key | Default | Range | Description |
|---|
microphone.vad.threshold | 0.5 | 0.3–1.0 | Speech probability threshold |
microphone.vad.min_silence_frame | 20 | 1–30 | Silence frames before segment end (× 10 ms) |
microphone.vad.min_speech_frame | 8 | 1–20 | Speech frames before segment start (× 10 ms) |
Shared Library
TEN VAD does not use an ONNX model. It requires the libten_vad shared library at both build and runtime.
Docker: The Dockerfile copies the library from the source tree:
COPY api/assistant-api/internal/vad/internal/ten_vad/lib/Linux/x64/libten_vad.so /opt/ten_vad/lib/
export CGO_CFLAGS="-I$(pwd)/api/assistant-api/internal/vad/internal/ten_vad"
export CGO_LDFLAGS="-L$(pwd)/api/assistant-api/internal/vad/internal/ten_vad/lib/Linux/x64 -lten_vad"
export LD_LIBRARY_PATH="$(pwd)/api/assistant-api/internal/vad/internal/ten_vad/lib/Linux/x64:$LD_LIBRARY_PATH"
The pre-compiled shared library is for Linux x86_64 only. TEN VAD is not available for macOS local development. Use Silero VAD or FireRed VAD instead when developing on macOS.
