Skip to main content
Silero VAD is a pre-trained ONNX model (~2 MB) for real-time voice activity detection. It is the default VAD provider in Rapida. Provider identifier: silero_vad

​
Source Location

api/assistant-api/internal/vad/internal/silero_vad/
├── silero_vad.go        # Main implementation
├── detector.go          # ONNX Runtime inference wrapper
├── models/
│   └── silero_vad_20251001.onnx   # Pre-trained model (~2 MB)

​
How It Works

  1. Incoming LINEAR16 bytes are converted to float32 samples in the range [-1.0, 1.0]
  2. Samples are fed to the Silero ONNX detector which produces speech segments with start/end timestamps
  3. On speech onset, an InterruptionPacket is emitted (triggers barge-in)
  4. While speech is active, VadSpeechActivityPacket heartbeats keep the EOS timer from firing
// Simplified flow
samples := linear16ToFloat32(pkt.Audio)
segments, isSpeaking, _ := detector.Detect(samples)

if hasSpeechStart(segments) {
    callback(InterruptionPacket{...})   // barge-in
}
if isSpeaking {
    callback(VadSpeechActivityPacket{}) // EOS timer reset
}

​
Parameters

Option KeyDefaultRangeDescription
microphone.vad.threshold0.50.3–1.0Speech probability threshold
microphone.vad.min_silence_frame201–30Silence frames before segment end (× 10 ms)
microphone.vad.min_speech_frame81–20Speech frames before segment start (× 10 ms)
Internally, frame counts are converted to milliseconds: min_silence_frame × 10 = MinSilenceDurationMs, min_speech_frame × 10 = SpeechPadMs.

​
Model Path

SourceResolution
Environment variableSILERO_MODEL_PATH — absolute path to the .onnx file
Default (Docker)./models/silero_vad/silero_vad_20251001.onnx
Default (source)api/assistant-api/internal/vad/internal/silero_vad/models/silero_vad_20251001.onnx

​
Local Source Setup

Silero VAD requires ONNX Runtime. The Docker base image includes it. For local builds: 
# macOS (Homebrew)
brew install onnxruntime

# Linux — download from https://github.com/microsoft/onnxruntime/releases
# Extract and set:
export CGO_CFLAGS="-I/opt/onnxruntime/include"
export CGO_LDFLAGS="-L/opt/onnxruntime/lib -lonnxruntime"
export LD_LIBRARY_PATH="/opt/onnxruntime/lib:$LD_LIBRARY_PATH"
The model file is checked into the repository — no download needed.