A free, open source, and extensible speech-to-text application that works completely offline.
//! The speech the language detector hears: silence cut out by VAD, so a
//! window of N seconds is N seconds of speech.
use anyhow::Result;
use crate::audio_toolkit::vad::{VadFrame, VoiceActivityDetector};
/// The first `max_samples` of speech in `audio`, as `vad` keeps it.
pub fn speech_head(
vad: &mut dyn VoiceActivityDetector,
audio: &[f32],
max_samples: usize,
) -> Result<Vec<f32>> {
let mut speech = Vec::with_capacity(max_samples.min(audio.len()));
for frame in audio.chunks_exact(vad.frame_samples()) {
if let VadFrame::Speech(kept) = vad.push_frame(frame)? {
speech.extend_from_slice(kept);
if speech.len() >= max_samples {
break;
}
}
}
speech.truncate(max_samples);
Ok(speech)
}
#[cfg(test)]
mod tests {
use super::*;
/// Frames whose first sample is positive are speech.
struct SignVad;
impl VoiceActivityDetector for SignVad {
fn push_frame<'a>(&'a mut self, frame: &'a [f32]) -> Result<VadFrame<'a>> {
Ok(if frame[0] > 0.0 {
VadFrame::Speech(frame)
} else {
VadFrame::Noise
})
}
fn frame_samples(&self) -> usize {
2
}
}
#[test]
fn keeps_only_speech_up_to_the_window() {
let audio = [0.0, 0.0, 1.0, 2.0, 0.0, 0.0, 3.0, 4.0, 5.0, 6.0];
assert_eq!(
speech_head(&mut SignVad, &audio, 3).unwrap(),
vec![1.0, 2.0, 3.0]
);
assert_eq!(
speech_head(&mut SignVad, &audio, 100).unwrap(),
vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]
);
assert!(speech_head(&mut SignVad, &[0.0; 6], 4).unwrap().is_empty());
}
}