What it does
Six seconds on the silence counter can release speech without a positive judgment. The maker tuned the first judge on one Chinese-speaking person; the second judge's cutoff remains untested.
Agent tooling
jev_voice_agent
A voice agent asks separately whether it has enough to answer and whether the person has finished speaking.
Bookmark: jev_voice_agent
Keep this in your collection.
Leave a noteWhat would you try with this?
: jev_voice_agent
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
How you can use it
For a talking assistant, separate preparing a reply from being allowed to speak it. Ask a developer to study how this project asks those two questions separately. Then test with your own speakers: the maker's tuning on one person does not establish that it suits other people or languages.
The live agent sends audio and text to paid outside services for speech recognition, judgments and reply writing. It also saves conversations on disk. Get participants' consent and set spending limits before trying it. Replacing the Jev judgments with fake answers does not turn off the other paid services.