Skip to main content
FireRed VAD is a DFSMN (Deep Feed-forward Sequential Memory Network) streaming model from the FireRed team at Ant Group. It uses Kaldi-compatible fbank feature extraction, CMVN normalization, and a 4-state postprocessor for precise speech boundary detection. Provider identifier: firered_vad

​
Source Location

api/assistant-api/internal/vad/internal/firered_vad/
├── firered_vad.go       # Main implementation
├── detector.go          # ONNX Runtime inference
├── fbank.go             # Kaldi-compatible fbank feature extraction
├── cmvn.go              # Cepstral mean and variance normalization
├── postprocessor.go     # 4-state speech boundary detector
├── models/
│   └── fireredvad_stream_vad_with_cache.onnx   # DFSMN model (~5 MB)

​
How It Works

  1. Incoming LINEAR16 bytes are converted to int16 samples and buffered
  2. Complete frames are extracted: 400 samples (25 ms) with 160-sample shift (10 ms)
  3. For each frame: fbank features are extracted, CMVN normalization is applied, ONNX inference produces a raw speech probability
  4. The postprocessor smooths probabilities (moving average, window size 5) and runs a 4-state machine:
Silence → Possible Speech → Speech → Possible Silence → Silence
  1. Speech onset is confirmed only after MinSpeechFrame consecutive frames above threshold — this filters out short noise bursts
  2. Same packet emission: InterruptionPacket on confirmed onset, VadSpeechActivityPacket heartbeats during confirmed speech

​
Parameters

Option KeyDefaultRangeDescription
microphone.vad.threshold0.50.3–1.0Speech probability threshold (applied after smoothing)
microphone.vad.min_silence_frame201–30Silence frames before segment end (each frame = 10 ms)
microphone.vad.min_speech_frame81–20Speech frames before confirming onset (each frame = 10 ms)

​
Internal Postprocessor Defaults

These are not configurable via assistant options — they are hardcoded in DefaultPostprocessorConfig():
SettingValueDescription
SmoothWindowSize5Moving average window for probability smoothing
SpeechThreshold0.4Smoothed probability threshold (separate from the configurable threshold)
PadStartFrame5Frames padded at speech start to capture onset audio (50 ms)
MaxSpeechFrame2000Max consecutive speech frames before forced reset (20 seconds)

​
Model Path

SourceResolution
Environment variableFIRERED_VAD_MODEL_PATH — absolute path to the .onnx file
Default (Docker)./models/firered_vad/fireredvad_stream_vad_with_cache.onnx
Default (source)api/assistant-api/internal/vad/internal/firered_vad/models/fireredvad_stream_vad_with_cache.onnx

​
Local Source Setup

FireRed VAD requires ONNX Runtime (same as Silero VAD). The model file is checked into the repository. 
# Ensure ONNX Runtime is available
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"