Block's AI learns language from 100% human audio without text
The company says the system, named Dessa, was built without text data and can run locally on specific Mac hardware.
By Yuna Fletcher ·
Block, the firm led by Jack Dorsey, announced an artificial intelligence model named Dessa that was trained exclusively on human conversation. The company states the system was built "from scratch on nothing but human conversation," a design choice that eliminates the use of written text during the training phase.
To demonstrate the model's capabilities, the team published audio recordings of exchanges with users. The results are mixed. In one sample shared by Ragavan Thurairatnam, who leads the project, the dialogue flows naturally, though the model initially misidentifies the speaker as "Ronaldo." Other clips on the company's website show the model taking unexpected tangents, such as discussing Christopher Nolan's films before shifting to Bryan Singer's work on Superman.
The developers describe Dessa as a "true speech-to-speech model," asserting that everything it knows is derived from audio input. They claim the system can operate locally on a MacBook Pro M4 Max or better hardware. This local execution capability distinguishes it from cloud-dependent voice assistants that often require internet connectivity.
The team outlines three primary goals for the project: curiosity, improved human-AI understanding, and safety. They compare the training method to child development, noting that babies learn to speak without seeing written words. The developers argue that current voice AI systems flatten speech into text and fail to match human timing, whereas Dessa is designed to handle audio directly.
A central part of the safety argument involves emotional recognition. The team contends that many AI systems remain "nearly deaf to the emotion in our voices." They cite a scenario where a person in pain interacts with a voice assistant that continues its standard responses without detecting the distress. The developers position Dessa as a step toward addressing this 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 or specific pricing details for the model.
Source: The Verge
