Block Releases Dessa, a Voice Model That Learned Without Text

Block announced Dessa, an artificial intelligence model trained solely on human audio, challenging the standard approach of converting speech to text. The company, led by Jack Dorsey, stated the system was built "from scratch on nothing but human conversation." This method aims to preserve the timing and emotional nuance of spoken language that text-based systems often lose.

The development team cites three goals for the project: curiosity, improved human-AI understanding, and safety. Ragavan Thurairatnam, who leads the project, compared the process to how babies learn to talk without seeing written words. The team noted that while the model can learn this way, the process is "very hard." They argue that current voice assistants "flattens what we say into text and can't keep up with human timing."

Block published several audio samples to demonstrate the model's capabilities. In one exchange shared by Thurairatnam, the conversation sounds mostly natural, though Dessa initially mistakes the speaker for someone named Ronaldo. Another sample shows the model discussing Christopher Nolan's films before drifting into a tangent about Bryan Singer's Superman movies. A third clip features a user telling the model about a surprise trip to Paris, prompting a response the site describes as "teary gratitude."

The team also released a sample pairing Dessa with OpenAI's GPT-Live-1 voice model. In that specific interaction, observers noted the model sounded more natural than in the standalone clips. The developers emphasize that speech is the most natural interface to intelligence and that their model is a "true speech-to-speech model, and everything it knows came from human speech."

Safety remains a primary driver for the release. The team argues that as AI systems gain power, they remain "nearly deaf to the emotion in our voices." They described a scenario where a person in serious pain speaks to a voice AI that carries on regardless. The team positions Dessa as "a step toward fixing" that disconnect by allowing the model to detect emotional cues in the audio.

The system is designed to run locally on hardware, specifically a MacBook Pro M4 Max or better. This local processing capability means the model does not require a constant cloud connection to function. The team did not provide a specific date for a wider public release but confirmed that early access is planned for a select group of experts.

Block plans to give "a small group of AI safety researchers" early access to the model. The announcement did not include a timeline for a general availability launch or additional technical specifications beyond the hardware requirements. The company has not yet disclosed how many hours of audio were used to train the system.