Voice activity detection: tuning it so it stops cutting me off

Mid-sentence pauses end my turn too early. What thresholds work?

Benchmarked on my machine: ~220 ms to first token with a quantised 7B.

Appreciate the honest feedback, even the harsh bits 😄

emre-y likes this.

Does VAD tuning work offline, or does it need the cloud? 🎉

Same!

@hannah-t Doing the same with VAD tuning, but with a local model, and it feels much more private.

Putri_S65, emre-y, tariqN and 11 others like this.

141

I tested VAD tuning on a cheap laptop and it held up better than expected 👍

The privacy side of VAD tuning worries me more than anyone here seems to admit.