Voice turn-end detection with Jev

Just a trigger. The demo waits half a second after speech stops, then asks Jev whether the turn really ended, and maps the returned score to extra hold time: the timeout=3.0 in the output is the added hold, and None means no hold added. The judgement is Jev's, not the silence timer's.

Mostly network, per the author. He attributes the long elapsed to latency from his own machine and expects a call hosted near the assistant to land in the 200 ms range. So read the logged figure as end-to-end, and subtract your own transport cost before drawing conclusions.

He thinks it might, saying the judgements are largely as expected, but he also says there is real tuning left to do. There are no accuracy numbers, no test set and no published code, so there is nothing to reproduce or compare against. Treat it as a promising prototype.

He raises it himself: since the cost is very low, calling Jev at the partial-recognition point rather than after the silence gate might work better. That is an untested idea in the thread, and it multiplies call volume, so measure before wiring it into a live pipeline.

うえぞう, posting as @uezochan, a hobby programmer who likes soba and serves as a Machida City digital promotion advisor and a LINE API Expert, and who maintains AIAvatarKit, a voice agent toolkit. So the turn-end check is a gate inside an existing voice-agent loop rather than a standalone experiment.

Japanese, and not reusable as published. The thread is a short text post with an attached screen recording and no repository or code, so the silence threshold, the score-to-hold mapping, and the state sent to Jev are all undocumented and would have to be rebuilt from scratch.

To avoid cutting a speaker off. When Jev scores the turn as incomplete, the gate adds hold time, shown as timeout=3.0, before the assistant takes its turn; None means it adds nothing. So the model is buying a pause when it judges the speaker is mid-thought rather than reclassifying the turn.