Block's new AI model learns language solely from human audio

Block, the company co-founded by Jack Dorsey, trained its newest artificial intelligence model on 100% human conversation. The system, named Dessa, was built without any written text. The team describes it as a speech-to-speech model that derives all its knowledge from audio.

The developers state that the model can operate locally on a MacBook Pro M4 Max or a more powerful machine. This local capability means the system does not need to send data to external servers to process voice. The project was driven by three main goals: curiosity, improved human-AI understanding, and safety.

Ragavan Thurairatnam leads the project at Block. He shared a sample recording where the model initially mistakes the speaker for a person named Ronaldo. The conversation then proceeds mostly naturally. Other examples on the company website show Dessa discussing Christopher Nolan films before shifting to a tangent about Bryan Singer's Superman movies.

In one recorded exchange, a user describes taking the model on a surprise trip to Paris. The website notes that Dessa responds with what sounds like teary gratitude. Another clip pairs Dessa with OpenAI's GPT-Live-1 voice model. In that specific setup, the developers say Dessa sounds more natural than in other demonstrations.

The team argues that current voice assistants reduce speech to text, which flattens the interaction. They claim this approach fails to match human timing. The safety argument focuses on emotion. The developers note that AI systems are gaining independence while remaining deaf to the emotional tone of the human voice.

They cite a scenario where a person in serious pain interacts with a voice AI that continues regardless because it cannot hear the distress. Block positions Dessa as a step toward fixing this gap. The company plans to give early access to a small group of AI safety researchers. No wider release schedule has been announced.

The developers compare the training process to babies learning to talk without seeing written words. They acknowledge that making this work is very hard. The model demonstrates that a system can learn language purely from sound, skipping the traditional text-based training phase entirely.