Clip / Other

Proposed fix: special silent tokens for turn-taking

From Why can't ChatGPT Voice set a timer? | Voice AI expert explains · ≈5:48

A concrete technical proposal for how audio codecs/token vocabularies could encode silence to let voice models yield the conversational floor properly.

What’s in it

A concrete technical proposal for how audio codecs/token vocabularies could encode silence to let voice models yield the conversational floor properly.

Watch this moment

Clip transcript

have the last word? >> You would probably want to do something like generating a silent response, or you have some special tokens to say this is, I don't know, empty tokens so that the audio codec knows to actually not use that token to generate any audio, and to wait for the next response