Block Launches Dessa Voice Model

Block's new AI can laugh in the middle of a sentence, sounding eerily human. The model, called Dessa, processes audio directly, so it can listen while it speaks. It was trained from scratch on a single Nvidia GB200 NVL72 GPU, using only human speech. The compact model can run on a consumer-grade MacBook Pro with an M4 Max chip. For now, Dessa remains a research preview, given only to a small group of AI safety researchers. Some users praised its natural cadence, while others warned it could make phone scams harder to detect. Block said the voice could eventually recognize emotional cues that text transcripts miss. "Talking with it feels like talking with a person...of emotion, nuance, pauses, and a natural cadence," Dorsey wrote on X.