Struct Endpointer 

pub struct Endpointer<V> { /* private fields */ }

Available on non-target_family=wasm only.

Finds utterances in a stream. Speech opens on a frame scored at or above SPEECH_THRESHOLD and closes once silence has lasted trailing_silence. After an End it keeps listening, so the next Start opens the next utterance. An utterance still open when input ends gets no End.

Positions count from the first push. To land them on a RecordingReader that did not start at zero, add its held_from() at that push. Start has no lead-in: open an utterance LEAD_IN_FRAMES frames earlier to cut as the microphone does.

fn utterances<V: VoiceActivityDetector>(
    vad: &mut V,
    batches: &[Vec<i16>],
) -> Result<(), VoiceActivityError> {
    // `&mut V` leaves the loaded detector with the caller for the next run.
    let mut endpointer = Endpointer::new(vad, SilenceDuration::default());
    for batch in batches {
        for boundary in endpointer.push(batch)? {
            match boundary {
                Boundary::Start(at) => println!("speech from sample {at}"),
                Boundary::End(at) => println!("utterance ends at sample {at}"),
            }
        }
    }
    Ok(())
}

Implementations

impl<V: VoiceActivityDetector> Endpointer<V>

pub fn new(vad: V, trailing_silence: SilenceDuration) -> Self

Resets vad, so state from earlier audio does not leak in.

pub fn push( &mut self, samples: &[i16], ) -> Result<Vec<Boundary>, VoiceActivityError>

A partial frame waits for the next push.