use loqui_speaker::{SAMPLE_RATE, SpeakerEncoder};
use crate::Error;
pub(crate) const MAX_EMBED_SECS: f32 = 30.0;
pub const MIN_EMBED_SECS: f32 = 1.0;
const FRAME: usize = SAMPLE_RATE as usize / 50;
#[derive(Debug, Clone, PartialEq)]
pub struct SpeakerEmbedding(Vec<f32>);
impl SpeakerEmbedding {
pub(crate) fn new(vector: Vec<f32>) -> Self {
Self(vector)
}
pub fn as_slice(&self) -> &[f32] {
&self.0
}
pub fn similarity(&self, other: &Self) -> f32 {
loqui_speaker::cosine(&self.0, &other.0)
}
}
pub(crate) fn voiced(samples: &[f32]) -> (&[f32], f32) {
let rms: Vec<f32> = samples.chunks(FRAME).map(|f| (f.iter().map(|s| s * s).sum::<f32>() / f.len() as f32).sqrt()).collect();
let peak = rms.iter().copied().fold(0.0f32, f32::max);
let threshold = (peak * 0.05).max(1e-4);
let loud: Vec<usize> = rms.iter().enumerate().filter(|(_, r)| **r >= threshold).map(|(i, _)| i).collect();
match (loud.first(), loud.last()) {
(Some(&first), Some(&last)) => {
let span = &samples[first * FRAME..((last + 1) * FRAME).min(samples.len())];
(span, loud.len() as f32 * FRAME as f32 / SAMPLE_RATE as f32)
}
_ => (&samples[..0], 0.0),
}
}
pub(crate) fn embed(encoder: &SpeakerEncoder, samples: &[f32], min_secs: f32) -> Result<SpeakerEmbedding, Error> {
let (span, secs) = voiced(samples);
if secs < min_secs {
return Err(Error::Invalid(format!(
"the recording holds {secs:.1} s of speech; describing a voice needs at least {min_secs:.0} s"
)));
}
let span = &span[..span.len().min((MAX_EMBED_SECS * SAMPLE_RATE as f32) as usize)];
Ok(SpeakerEmbedding::new(encoder.embed(span)?))
}
#[cfg(test)]
mod tests {
use super::*;
fn tone(secs: f32, amplitude: f32) -> Vec<f32> {
(0..(secs * SAMPLE_RATE as f32) as usize).map(|i| amplitude * (i as f32 * 0.1).sin()).collect()
}
#[test]
fn silence_at_either_end_is_trimmed() {
let mut samples = vec![0.0; SAMPLE_RATE as usize];
samples.extend(tone(2.0, 0.3));
samples.extend(vec![0.0; SAMPLE_RATE as usize / 2]);
let (span, secs) = voiced(&samples);
assert!((span.len() as f32 / SAMPLE_RATE as f32 - 2.0).abs() < 0.03, "{}", span.len());
assert!((secs - 2.0).abs() < 0.03, "{secs}");
}
#[test]
fn a_pause_inside_is_kept_but_not_counted() {
let mut samples = tone(1.0, 0.3);
samples.extend(vec![0.0; SAMPLE_RATE as usize * 3]);
samples.extend(tone(1.0, 0.3));
let (span, secs) = voiced(&samples);
assert!((span.len() as f32 / SAMPLE_RATE as f32 - 5.0).abs() < 0.03);
assert!((secs - 2.0).abs() < 0.05, "{secs}");
}
#[test]
fn digital_silence_holds_no_speech() {
assert_eq!(voiced(&[0.0; 32_000]).1, 0.0);
assert_eq!(voiced(&[]).1, 0.0);
}
#[test]
fn similarity_is_the_cosine() {
let a = SpeakerEmbedding::new(vec![1.0, 0.0]);
let b = SpeakerEmbedding::new(vec![0.0, 1.0]);
assert!((a.similarity(&a) - 1.0).abs() < 1e-6);
assert!(a.similarity(&b).abs() < 1e-6);
}
}