Voxta LiveKit Turn Detector
A small local model that reads the transcript to judge whether you are finished.
Runs LiveKit's distilled end-of-utterance model locally on the CPU via ONNX Runtime. It reads the transcript rather than its punctuation, which is a much better judge than cue lists.
Unlike the built-in module, this one replaces your transcription service's Complete and Incomplete Sentence Timeouts while it is active.
Setup
Add the service
Manage Services → + Add Services → LiveKit Turn Detector → Add. The model downloads on first use.
Pick a model
| Model | Covers | Notes |
|---|---|---|
| English (default) | English | Small and fast. |
| Multilingual | 14 languages | Considerably larger and slower. |
Set the silence range
Silence Before Replying (ms) is a range with two ends:
- Model says finished — the low end. Shorter feels snappier; too short clips the tail of your last word.
- Model says mid-sentence — the high end, and also how long you have to carry on into the same turn.
Everything between is interpolated, weighted so that only genuinely unfinished-sounding transcripts reach out toward the long end.
Needs an ONNX Runtime build for your platform. Where there isn't one, Voxta falls back to the built-in detector rather than losing the feature.