Voxta Turn Detection
Deciding when you have actually finished speaking, rather than just gone quiet.
No single silence threshold is right. The same 600 ms means "done" after "what's your name?" and "still going" after "I was thinking that maybe we could—". A short threshold cuts you off whenever you pause to think; a long one makes every exchange sluggish.
Turn detection judges how finished your sentence sounds and scales the wait to match. It sits under Speech Transcription → Turn Detection.
This costs nothing when you are genuinely finished. The extra wait is only ever spent on a transcript that sounds unfinished — and it is also the window you have to keep talking into the same turn.
Providers
Voxta Utilities
Punctuation and trailing words. No model, no download, on by default.
LiveKit Turn Detector
A small local model that reads the transcript. More accurate.
Choosing None gives you back plain silence timeouts — your transcription service's Complete and Incomplete Sentence Timeouts, and nothing else.
The difference between them
The built-in module adds to whatever your transcription service was already going to wait, and only when it sees a positive cue that you are mid-sentence. It is trustworthy in one direction: a trailing "and" really does mean you are not done, while a trailing period is just the recognizer guessing.
The LiveKit module replaces that timing. A model reading the whole transcript judges it far better than punctuation does, so blending the two would only drag a good answer back toward a worse one.
Watching it work
The Speech-to-Text playground shows turn detection live, and lets you toggle it off to compare.