silero_vad

Silero VAD for ExecuTorch

Export and run Silero VAD (16kHz, ~2MB, MIT license) on ExecuTorch for native C++ voice activity detection.

Voice activity detection answers "when is someone speaking" — the model outputs a speech probability (0.0–1.0) for each 32ms audio chunk.

Quick Start

# Export to .pte
cd examples/models/silero_vad
python export_silero_vad.py \
    --jit-model /path/to/silero_vad.jit \
    --backend xnnpack \
    --output-dir ./silero_vad_xnnpack

# Build the C++ runners (from repo root)
make silero-vad-cpu

# Run WAV-based VAD
./cmake-out/examples/models/silero_vad/silero_vad_runner \
    --model_path examples/models/silero_vad/silero_vad_xnnpack/silero_vad.pte \
    --audio_path /path/to/audio.wav

Output:

  12.512 12.672 speech
  12.736 13.216 speech
  13.248 13.696 speech
  13.728 13.952 speech
  13.984 14.144 speech
  14.176 14.240 speech
  14.272 14.432 speech
  18.144 18.176 speech
  19.168 19.200 speech
  22.624 23.776 speech
  25.216 25.312 speech
  25.344 26.240 speech
  26.304 26.592 speech
  29.376 30.464 speech
  ...

502 segments, 32793 frames, 1049.4s
Speech: 19489/32793 frames (59.4%)

Each line is start_seconds end_seconds speech. Each output frame covers 32ms of audio (512 samples at 16kHz).

Getting the Model

The export script requires the silero_vad.jit file from the Silero VAD repository:

git clone https://github.com/snakers4/silero-vad.git

The JIT model is at silero-vad/src/silero_vad/data/silero_vad.jit.

Export

python export_silero_vad.py --jit-model /path/to/silero-vad/src/silero_vad/data/silero_vad.jit

Argument	Description
--jit-model	Path to silero_vad.jit file (required)
--backend	portable or xnnpack (default: xnnpack)
--output-dir	Output directory (default: ./silero_vad_exports)

Output: silero_vad_xnnpack/silero_vad.pte (~1.2 MB with XNNPACK, may vary by export settings).

C++ Runners

Build

From the repository root:

make silero-vad-cpu

This builds:

• cmake-out/examples/models/silero_vad/silero_vad_runner
• cmake-out/examples/models/silero_vad/silero_vad_stream_runner

The build configures and installs ExecuTorch first, then builds the two Silero VAD binaries from examples/models/silero_vad/.

Arguments

silero_vad_runner

Argument	Description
--model_path	Path to .pte file (default: silero_vad.pte)
--audio_path	Path to input WAV file (16kHz mono, required)
--threshold	Speech probability threshold, 0.0–1.0 (default: 0.5)

silero_vad_stream_runner

Argument	Description
--model_path	Path to .pte file (default: silero_vad.pte)

The stream runner reads 16kHz mono float32 PCM from stdin and prints:

READY
PROB <time_seconds> <probability>

Example:

ffmpeg -i input.wav -ar 16000 -ac 1 -f f32le -nostats -loglevel error pipe:1 | \
  ./cmake-out/examples/models/silero_vad/silero_vad_stream_runner \
    --model_path examples/models/silero_vad/silero_vad_xnnpack/silero_vad.pte

How It Works

The model processes audio in 512-sample chunks (32ms at 16kHz). Each chunk is prepended with 64 samples of context from the previous chunk, forming a 576-sample input. The model carries an LSTM hidden state across chunks and outputs a single speech probability per chunk.

The runner applies a simple threshold to produce speech segments: consecutive frames above the threshold are merged into a single segment.

Limitations

• Input must be 16kHz mono WAV.
• No smoothing or minimum segment duration filtering — raw threshold output.

Architecture

Input: audio (1, 576), LSTM state (2, 1, 128)

Learned STFT:  Conv1d(1, 258, 256, stride=128)  →  magnitude spectrum (1, 129, 4)
CNN encoder:   Conv1d(129→128→64→64→128, kernel=3) →  (1, 128)
LSTM:          LSTMCell(128, 128)                 →  h, c
Decoder:       Conv1d(128, 1, 1) → Sigmoid        →  probability (1, 1)

Output: speech probability (1, 1), new LSTM state (2, 1, 128)

Exported Method

The .pte contains a single method:

Method	Backend	Input	Output
forward	XNNPACK	x (1, 576) float, state (2, 1, 128) float	prob (1, 1) float, new_state (2, 1, 128) float

Metadata via constant_methods: sample_rate (16000), window_size (512), context_size (64).