Yeah, it’s getting harder to tell. At some point the difference won’t be in the voice itself, but in how the conversation flows.

Even now, i think they're quick enough--but they interrupt at the wrong times, where humans know if they have enough context yet.

So, I agree. But I believe the problem is pretty solvable with enough tokens.

Impressive. Running a 400B model on-device, even at low throughput, is pretty wild.