flexaudio-vad

Offline Voice Activity Detection for Rust, powered by the Silero VAD model running on ONNX Runtime. The model is embedded into the binary via include_bytes!, so detection runs fully offline — no model file to ship and no network access at runtime.

This crate is independent of flexaudio-core: it consumes a plain &[f32] sample stream, so you can pair it with any audio source.

Streaming example

use flexaudio_vad::{Vad, VadConfig, VadEvent};

let mut vad = Vad::new(VadConfig::default()).unwrap();
for chunk in some_audio_chunks() {
    for ev in vad.process(chunk) {
        match ev {
            VadEvent::SpeechStart { at_sample } => println!("start @ {at_sample}"),
            VadEvent::SpeechEnd { at_sample }   => println!("end @ {at_sample}"),
        }
    }
}
# fn some_audio_chunks() -> Vec<&'static [f32]> { vec![] }

Batch example

use flexaudio_vad::{get_speech_timestamps, VadConfig};

let samples: Vec<f32> = vec![0.0; 16_000];
let segments = get_speech_timestamps(&samples, &VadConfig::default()).unwrap();
for s in segments {
    println!("{}..{}", s.start_sample, s.end_sample);
}

VadConfig ships aggressive(), balanced(), and conservative() presets. Input is expected as 16 kHz mono f32.

Install

cargo add flexaudio-vad

MSRV

Rust 1.88 (required by the ort / ONNX Runtime toolchain).

License & third-party notices