Trained on speech alone, no text: Jack Dorsey's Block unveils Dessa, a voice AI that laughs and cuts in

A voice AI that learned to talk only by listening to human conversation is here. Dessa, released Oct 23 by a small AI lab at Jack Dorsey's Block, never saw text.

Block says Dessa was trained from scratch on human speech, with no text and no pretrained models. It is full-duplex, listening while it talks, so it laughs mid-sentence, backchannels with "yeah" and "got it," and speaks louder when asked. Block trained it on a single Nvidia GB200 NVL72 and says it runs locally on a MacBook Pro with an M4 Max or better. It learned language from hundreds of times fewer words than text LLMs, the company said.

"Talking with it feels like talking with a person," Dorsey wrote, in a post with more than 19,000 likes. Lab lead Ragavan Thurairatnam called it "a GPT3 moment for speech." One demo has Dessa talking over and with ChatGPT's GPT-Live-1.

Replies leaned into jokes: "Grandmas of the world are COOKED" drew 318 likes, and "What did you train this on, all my friends from Queens?" drew 97. Block is inviting a small group of AI safety researchers to probe the model's risks.