Voxta Voxta Utilities: Turn Detection
Punctuation and trailing-word cues. No model, no download, enabled by default.
The built-in turn detector reads punctuation and trailing words to tell whether you sound mid-sentence. There is no model and nothing to download, which is why it ships enabled.
It only ever adds to the timeouts your speech transcription service already uses, and only on a positive cue — a trailing comma, a conjunction, a verb ending that implies more is coming. A transcript with no clear cue waits exactly what it would have with the module absent.
Setting
Extra Silence When Unfinished (ms) — the most it will add on top of your transcription module's own timeouts. Default is a few hundred milliseconds; the range goes to 6000.
Being generous costs nothing when you are actually finished, because the wait is only spent on someone who trailed off. It is also how long you have to carry on speaking into the same turn.
Set it to 0 and the module is off in practice.
Coverage
Cue lists ship for English, Spanish, French, German, Portuguese, Japanese, Chinese and Korean. Japanese and Korean are matched on verb endings rather than word lists, since that is the shape their morphology takes.
The probabilities behind the cues are measured against LiveKit's public end-of-turn benchmark, not hand-picked — about 7,800 labelled judgement points.
It stands aside automatically if you select another turn detection module, and picks the job back up if that one cannot run.