Crossing the uncanny valley of conversational voice

https://www.sesame.com/research/crossing_the_uncanny_valley_of_voice

Conversational Speech Model: Llama-variant backbone (1B, 3B, 8B) plus small decoder over Mimi split-RVQ tokens at 12.5 Hz; about one million hours of mostly English audio. CMOS: no preference vs human without context; human recordings preferred with 90 s of context. No turn-taking modeling; Apache 2.0 release announced.

Realtime speech-to-speech and expressive TTS for characters