Struct Chunker 

pub struct Chunker {
    pub samples_since_last_stride: usize,
    pub total_samples_seen: u64,
    pub last_speech_at_sample: Option<u64>,
    /* private fields */
}

Fields

samples_since_last_stride: usize

total_samples_seen: u64

last_speech_at_sample: Option<u64>

Implementations

impl Chunker

pub fn new(cfg: WindowConfig, vad: SileroVad) -> Self

pub fn push(&mut self, samples: &[f32]) -> Option<StreamWindow>

Push new microphone samples; returns Some(window) when a stride boundary fires or the buffer hits the max_window_secs cap.

The buffer GROWS — it doesn’t slide. Consumers must call [trim_oldest] or [reset_utterance] after a cap-hit window so the next push has room.

On cap-hit, the call returns early with the cap-hit window; any remaining input samples are deferred internally and prepended to the next push call. This avoids the failure mode where multiple cap-hits fire within one batch and only the last (forced_eou) window reaches the consumer.

pub fn reset_utterance(&mut self)

Drop the current utterance buffer so the next stride starts fresh. Call after committer.finalize_utterance() (either natural EOU or forced_eou). Wall-clock counters and the last-speech timestamp persist so trailing-silence math stays valid across utterance boundaries.

pub fn trim_oldest(&mut self, samples: usize) -> usize

Drop the oldest samples from the front of the buffer at a committed-token boundary, keeping the utterance going. Rounded down to a VAD_FRAME_SIZE multiple so vad_decisions stays in lock-step with buf. Advances utterance_start_sample to preserve wall-clock anchoring; current_secs = window_start_secs + real_samples/SR is invariant (the increase in window_start exactly cancels the decrease in real_samples).

Returns the number of samples actually trimmed (0 if the requested amount is less than one VAD frame or would empty the buffer). Returning 0 leaves cap_pending_handler set so the next cap escalates to forced-EOU.