svad
Voice activity detection using Silero VAD v6.2 weights. Supports mono 16 kHz and 8 kHz audio, speech probabilities, and speech segmentation.
Build
To build the library, you need Make, a C11 compiler (GCC or Clang), and ar.
Get the weights
This step needs curl, Python 3, and uv. The uv command below installs NumPy and ONNX for the script. Python is not needed to compile or run the C library.
Run these commands from the project folder. They download the Silero VAD v6.2 models and create weights_16k.h and weights_8k.h. Both headers must be in this folder before you run make:
mkdir -p build/models
base=https://raw.githubusercontent.com/snakers4/silero-vad/v6.2/src/silero_vad/data
curl -fL "$base/silero_vad_16k_op15.onnx" -o build/models/silero_vad_16k_op15.onnx
curl -fL "$base/silero_vad.onnx" -o build/models/silero_vad.onnx
uv run --no-project --with numpy --with onnx python scripts/export_weights.py \
build/models/silero_vad_16k_op15.onnx \
build/models/silero_vad.onnx \
build/generated
cp build/generated/src/weights_16k.h build/generated/src/weights_8k.h .
The script also creates build/generated/svad_weights.bin. You only need that file if you use svad_create_from_file to load weights at run time.
Build the library and command
make # builds libsvad.a and svad
make debug # optional debug build
make clean # removes build outputs
Link your application with libsvad.a and -lm.
Command-line usage
svad reads a 16-bit PCM WAV file at 8 kHz or 16 kHz. Multichannel audio is downmixed to mono. It prints speech start/end times in seconds and sample positions.
./svad audio.wav
./svad --json --quiet audio.wav > segments.json
./svad --stream audio.wav
Usage:
svad [--stream] [--probs FILE] [--json] [--quiet]
[--threshold X] [--neg-threshold X] [--min-speech-ms N]
[--max-speech-s X] [--min-silence-ms N] [--speech-pad-ms N]
[--weights FILE] [--scalar] [--dense-stft] [--no-ftz] INPUT.wav
--streamprints speech start/end events while processing the file.--jsonformats offline segments as JSON.--quietsuppresses statistics.--probs FILEsaves one native float32 probability per 32 ms chunk.
Use --help to display the available options.
Library usage
Save this example as example.c. It reads raw mono signed 16-bit PCM in native byte order from standard input at 16 kHz and prints one speech probability per 32 ms chunk. Input must already be decoded and resampled; WAV headers are not accepted.
#include "svad.h"
#include <stdio.h>
int main(void)
{
svad_t *vad = svad_create(16000);
if (!vad) return 1;
const size_t chunk_size = (size_t)svad_chunk_samples(vad);
int16_t chunk[512];
size_t n;
while ((n = fread(chunk, sizeof(*chunk), chunk_size, stdin)) > 0) {
for (size_t i = n; i < chunk_size; ++i) chunk[i] = 0;
printf("%.6f\n", svad_process_s16(vad, chunk));
}
svad_destroy(vad);
return ferror(stdin) ? 1 : 0;
}
cc -std=c11 -I. example.c libsvad.a -lm -o example
./example < audio.pcm
Use svad_create(8000) for 8 kHz audio. Each call consumes exactly 512 samples at 16 kHz or 256 at 8 kHz, without overlap. svad_process_f32 accepts float samples in [-1, 1]. Call svad_reset before starting a new audio stream and use a separate instance for each concurrent stream.
For speech boundaries, feed probabilities to svad_segmenter_update after svad_segmenter_init, or use svad_speech_timestamps for a complete recording. See svad.h for the API and segmentation parameters.
