Block's AI Model Trained on 100% Human Audio
The company led by Jack Dorsey says its new system, Dessa, learned language without reading a single written word.
By Luana Foster ·
Block has released an artificial intelligence model called Dessa that was trained exclusively on spoken words. The company stated the system was built from scratch using nothing but human conversation. It does not rely on written text to learn language.
Ragavan Thurairatnam, who leads the project at Block, shared sample recordings of the model interacting with people. In one exchange, the model initially mistakes the speaker for a person named Ronaldo. The conversation otherwise flows naturally. Other examples on the project website show the model discussing Christopher Nolan films before drifting into a discussion of Bryan Singer's Superman movies.
In another recording, a user tells the model about a surprise trip to Paris. The site describes Dessa's response as sounding like teary gratitude. A separate clip pairs Dessa with OpenAI's GPT-Live-1 voice model. In that specific interaction, the model's speech patterns appear more natural.
The development team describes the system as a true speech-to-speech model. They claim it can run locally on a MacBook Pro M4 Max or better hardware. The team listed three main goals for the project: curiosity, better understanding between humans and AI, and safety.
Regarding the curiosity goal, the developers compared the process to how babies learn to talk without seeing written words. They noted that while it is possible for a model to learn this way, the process is very difficult. The team also argued that current voice assistants flatten human speech into text. This approach, they said, fails to keep up with human timing.
The safety argument focuses on emotion. The team stated that current AI systems are nearly deaf to the emotion in human voices. They described a scenario where a person in serious pain speaks to an AI that carries on regardless. They positioned Dessa as a step toward fixing that issue.
Block plans to give early access to a small group of AI safety researchers. The announcement did not include a timeline for a wider public release.
Source: The Verge
