🎧 A small research lab Block showed Dessa, the most realistic voice AI model yet, which learned to speak from sound alone.
The conversation feels like talking to a real person: with emotions, pauses and natural rhythm. It listens and speaks at the same time, so it interrupts at the right moment, backchannels ("uh-huh", "mhm", "ooh") and laughs mid-sentence when something is funny.
Regular voice assistants first convert speech into text and lose intonation, emotion, sarcasm, sighs and voice tremors. Dessa works on a "sound in, sound out" scheme and perceives all of this directly.
It was trained only on human conversations: without text and without other pre-trained models. This is the world's first AI for live conversation that learned exclusively from sound.
It runs faster than real time and runs locally on a MacBook Pro with M4 Max or a more powerful model.
