Block releases Dessa, a speech-to-speech AI trained solely on human conversation

Block, led by Jack Dorsey, released Dessa, a speech-to-speech AI model trained “from scratch on nothing but human conversation.” The team says Dessa is a “true speech-to-speech model,” meaning it avoids flattening speech into text and instead processes voice directly. Block provided audio examples showing Dessa engaging in back-and-forth conversations with humans, including a mistaken greeting (“Ronaldo”) and a tangent about Bryan Singer’s Superman films during a discussion of Christopher Nolan’s work. The model can run locally on a MacBook Pro M4 Max or better, according to the team.

The Dessa team frames the model as a response to perceived gaps in current voice AI, arguing that today’s systems waste the richness of speech by converting it to text and struggle with human timing and emotional cues. They cite safety concerns, such as an AI failing to detect a user’s pain because it cannot interpret vocal emotion. Block plans to give “a small group of AI safety researchers” early access to Dessa, positioning the release as both a technical demonstration and a contribution to safety research.

The announcement includes audio clips and videos on the Dessa website, where Block highlights examples such as a user telling Dessa about a surprise trip to Paris and Dessa responding “with what sounds like teary gratitude.” The team also shared a clip of Dessa interacting with OpenAI’s GPT-Live-1 voice model, claiming Dessa came across as more natural in that exchange.