Block releases AI model that learns language only from spoken words

100 percent of the training data for Block's new artificial intelligence model came from human conversation. The company, led by Jack Dorsey, announced the release of Dessa on October 23, 2026. Block stated that the system was built "from scratch on nothing but human conversation." The model represents a shift away from standard voice assistants that convert speech to text. Dessa processes audio directly to generate spoken responses.

Block has not provided a detailed timeline for how the model was developed. The team identified three main reasons for the project. They cited curiosity, a desire for better human-AI understanding, and safety improvements. The developers noted that babies learn to talk without seeing written words. They wanted to test if an AI could do the same. The team described the process as "very hard" but achievable. They also argued that current voice systems flatten human speech into text. This method, they said, fails to match human timing.

Block published several audio samples to demonstrate the model's capabilities. Ragavan Thurairatnam, who leads the Dessa project, shared one exchange. In that clip, the model initially mistook the speaker for a person named Ronaldo. The rest of the conversation sounded mostly natural. The company website hosts additional examples. One sample features a discussion about Christopher Nolan films that drifts into a tangent about Bryan Singer's Superman movies. Another clip involves a user sharing a surprise trip to Paris. The site notes that Dessa responds with what sounds like teary gratitude. A final sample pairs Dessa with OpenAI's GPT-Live-1 voice model, where the interaction appears more natural.

The developers emphasize that speech is the most natural interface for intelligence. They argue that existing voice AI often lacks emotional awareness. The team described a scenario where a person is in serious pain but the AI carries on regardless. They positioned Dessa as a step toward fixing this issue. The system is designed to detect emotion in the user's voice. This focus on emotional data is central to their safety argument. They believe AI systems are gaining power while remaining deaf to vocal cues.

Dessa is a true speech-to-speech model. The team stated that everything the system knows came from human speech. It does not rely on text data for training. The model can run locally on a MacBook Pro M4 Max or better hardware. This local capability allows the system to operate without constant cloud connectivity. Block plans to provide early access to a small group of AI safety researchers. The announcement did not include a schedule for a wider public release.