Collect Speech Turn

@bleenco1 VersionUtility

Buffers an AudioFrame stream guided by VAD spans ([{start: Double, end: Double}]) and a turn-end Double probability, emitting one padded utterance AudioFrame per completed speech turn plus a Bool collecting flag.

How it fits

Typical backends

Detect voice activity on incoming audio, buffer each turn, and transcribe the flushed utterance.

Use a dedicated end-of-turn detector for the turn-end probability and transcribe each completed utterance.

Caveats

I/O contractThe utterance output emits only when a turn completes, while the collecting flag is produced every tick; downstream stages reading the utterance see bursts, not a per-tick stream.
State lifetimeAudio is buffered in memory across ticks for the current turn only and flushed on turn end, so the buffer never grows beyond one utterance.
Parameter interactionA turn flushes when the span list empties and the turn-end probability clears the threshold; an utterance shorter than the minimum duration is dropped, and one growing past the maximum is force-flushed early.
I/O contractLeading and trailing padding windows are kept around the detected speech so onset consonants and trailing sounds are not clipped from the emitted clip.

Versions

6ed8a48dlatestdefault
Automated release
6/21/2026