Block releases speech-only AI model trained entirely on human conversation

Block, the payment and technology company led by Jack Dorsey, announced the release of Dessa, an artificial intelligence model trained exclusively on spoken human audio. The company stated the system was built from scratch without any written text data. To demonstrate the model's capabilities, Block published audio clips of interactions between Dessa and various individuals. Some of these exchanges appear natural, while others sound awkward or uncanny.

Ragavan Thurairatnam, who leads the project, shared one sample interaction. In that recording, the conversation flows smoothly for the most part, although Dessa initially mistakes the speaker for a person named Ronaldo. The company's website hosts additional examples that showcase the model's varying degrees of success. In one clip, a discussion about Christopher Nolan's films drifts into an unexpected topic regarding Bryan Singer's Superman movies. In another, a user describes a surprise trip to Paris, prompting a response from Dessa that the team describes as sounding like teary gratitude. A third clip pairs Dessa with OpenAI's GPT-Live-1 voice model, where the speech patterns sound more natural than in the standalone examples.

The development team describes Dessa as a true speech-to-speech model where every piece of knowledge derives from human speech. They confirmed the system can operate locally on a MacBook Pro M4 Max or a machine with superior specifications. The team outlined three primary goals for the project: curiosity, improved human-AI understanding, and safety. They compared the approach to child development, noting that babies learn to speak without ever seeing written words. The team acknowledged that while the model achieved this goal, the process remained very difficult.

Developers argued that current voice assistants underutilize the potential of spoken language. They contended that speech is the most natural interface for intelligence, but existing systems flatten spoken words into text and fail to match human timing. The safety aspect of the project focuses on emotional recognition. The team noted that AI systems are gaining power and independence while remaining nearly deaf to vocal emotion. They described a scenario where a person in serious pain interacts with a voice assistant that continues as if nothing is wrong. The team labeled Dessa a step toward fixing that gap.

Block plans to provide early access to a small group of AI safety researchers. The announcement did not include a timeline for a wider public release.