Jack Dorsey's company releases voice model that learns from speech alone

Block, the firm led by Jack Dorsey, has introduced an artificial intelligence system named Dessa.

The company stated that the model was built from scratch using only human conversation. The developers describe it as a true speech-to-speech system. They claim everything the model knows originated from audio. The team noted that the software can operate locally on a MacBook Pro M4 Max or similar hardware.

The project team outlined three primary goals for the work. These include curiosity, improving human AI interaction, and safety. The developers compared the training process to infant development. They observed that babies learn to speak without seeing written words. The team said they wanted to test if a machine could do the same. They noted the task is very difficult but possible.

Ragavan Thurairatnam, who leads the project, shared a sample interaction on social media. In the recording, the model initially mistakes the speaker for a person named Ronaldo. The rest of the exchange sounds natural. The company website features additional clips. One conversation shifts from Christopher Nolan films to Bryan Singer movies. Another clip involves a user describing a surprise trip to Paris. The website describes the model's response as teary gratitude. A separate recording pairs Dessa with OpenAI's GPT-Live-1 voice model. In that specific test, the output sounds more natural.

The developers argue that current voice assistants limit the potential of speech. They state that speech is the most natural interface for intelligence. They criticize existing tools for converting speech to text and failing to match human timing. The team also highlighted a safety concern regarding emotion. They noted that AI systems are gaining independence while remaining deaf to vocal emotion. They described a scenario where a person in pain is ignored by a voice assistant. They positioned Dessa as a step toward correcting this issue.

Block plans to provide early access to the model for a small group of AI safety researchers. The announcement did not include a timeline for a wider release.