Reset & stream boundaries

Wake-word detection is stateful. The frontend retains context and the detector retains model input rows and recent probabilities across calls.

Call reset() when continuity is broken:

detector.reset()?;

Reset after

• dropped or missing audio blocks;
• seeking within a file;
• switching callers, speakers, tracks, or input devices;
• restarting a decoder;
• starting an unrelated recording;
• a network discontinuity.

Without a reset, the beginning of one source may be evaluated using feature history from another source. That can distort probabilities or create false activations near the boundary.

Do not reset between normal blocks

// Correct: state flows across contiguous blocks.
for block in contiguous_audio.chunks_exact(AUDIO_BLOCK_SAMPLES) {
    detector.process_audio(block)?;
}

// Boundary: clear state once before an unrelated stream.
detector.reset()?;

Listener manages this automatically when it detects an internal queue overflow. Low-level Detector callers own this responsibility because only they know whether their input remains continuous.