Block's Dessa Learns Speech From Human Talk, Runs on MacBook Pro M4 Max

Block has built a voice model called Dessa that learned to talk the way a child does: only by listening.

The company, led by Jack Dorsey, says Dessa was trained from scratch on human conversation, with no written text fed into it. Ragavan Thurairatnam, who leads the Dessa project, posted a recording of himself talking with the model. In that exchange, the conversation sounds mostly natural, though Dessa first mistakes Thurairatnam for someone named Ronaldo.

The Dessa website carries several more recordings. One drifts from Christopher Nolan's films into a tangent about Bryan Singer's Superman movies. Another has a person telling Dessa about a surprise trip to Paris, where the site says the model then answers with what sounds like teary gratitude. A third clip pairs Dessa with OpenAI's GPT-Live-1 voice model, and in that exchange Dessa sounds more natural.

The Dessa team describes the system as a true speech-to-speech model and says everything it knows came from human speech. It can run on a MacBook Pro M4 Max or better, without sending audio to a remote server.

The team lists three reasons for building it: curiosity, better understanding between humans and AI, and safety. On the first, it compares the model to an infant, saying babies learn to talk without ever seeing a word written down and that it wanted to see if a model could too. The developers also say current voice assistants flatten speech into text and cannot keep up with human timing.

The safety argument turns on emotion. The team says AI systems grow more powerful and independent while staying nearly deaf to the emotion in our voices, and it offers the example of a person in serious pain whose voice AI has no idea and carries on regardless. Block plans to give a small group of AI safety researchers early access to Dessa. The announcement did not set out a wider release schedule.