Block unveils Dessa, a speech model trained only on human conversation

Block's small AI research lab has released Dessa, a full-duplex speech-to-speech model built by former members of the Toronto startup Dessa. According to Block, the model was trained from scratch on raw human conversation audio only, with no text and no pretrained models involved, and it listens and talks at the same time rather than taking turns like most voice assistants.

Block says Dessa picks up behaviors nobody explicitly taught it, such as laughing mid-sentence, getting louder when asked to speak up, and using backchannel sounds like "yeah" or "got it" to show it is following along. The company says the model runs faster than real time and can run locally on a MacBook Pro M4 Max or better, and that it was trained using a single NVIDIA GB200 NVL72 system. These are Block's own claims, not independently verified.

Jack Dorsey, Block's cofounder and chairman, posted on X that voice is the most natural way to interact with intelligence and that talking with Dessa feels like talking with a person, full of emotion, nuance, pauses and natural cadence. Block says it is inviting outside AI safety researchers to get early access and probe the model for risks the team hasn't anticipated.

Why it matters

It's a rare attempt to build a voice AI that understands emotion and timing directly from speech instead of flattening conversations into text first.

Source: Block