Block releases AI model trained entirely on spoken audio
The system, named Dessa, was built without text data and runs locally on high-end Mac hardware.
By Layla Fontaine ·
Block has launched a new artificial intelligence model that learned language without processing a single written word. The company led by Jack Dorsey announced the release of Dessa, a system trained from scratch on nothing but human conversation. This approach marks a departure from standard models that rely heavily on written text for their foundational knowledge.
The development team led the project with three stated goals: curiosity, improved human-AI interaction, and safety. Ragavan Thurairatnam, who leads the initiative, compared the learning process to infants. He noted that babies learn to talk without seeing written words. The team aimed to determine if a machine could replicate that specific cognitive path. They confirmed it can, though they described the engineering challenge as very hard.
Block published several audio recordings to demonstrate the model's capabilities. In one clip shared by Thurairatnam, the exchange appears mostly natural, although the model initially mistakes the speaker for a person named Ronaldo. Another sample on the project website shows the AI discussing Christopher Nolan films before drifting into a tangent about Bryan Singer's Superman movies. A third recording features a user describing a surprise trip to Paris. The site notes that Dessa responds with what sounds like teary gratitude in that segment.
The company also provided a demonstration pairing Dessa with OpenAI's GPT-Live-1 voice model. In that specific exchange, the model's output sounds more natural than in the standalone clips. Block states that the system functions as a true speech-to-speech model. All of its knowledge was derived directly from human speech patterns.
The developers argue that current voice assistants limit the potential of spoken interaction. They claim that existing technology flattens human speech into text and fails to match human timing. The team also emphasized the safety aspect of the new model. They stated that AI systems are gaining power while remaining nearly deaf to the emotion in human voices. They cited a scenario where a person in pain interacts with a voice AI that carries on regardless. Block described Dessa as a step toward fixing that disconnect.
The hardware requirements for running the system are specific. The team says Dessa can run locally on a MacBook Pro M4 Max or better. This allows the model to function without sending data to external servers. Block plans to grant early access to a small group of AI safety researchers. The announcement did not include a schedule for a wider public release.
