Block unveils Dessa, a full-duplex voice model trained only on human speech
The model runs locally, laughs and pauses, and is being offered first to AI security researchers for testing rather than to the public.
Block, the company behind Square and Cash App, has unveiled Dessa, a speech model from its research lab. Block says it is the first AI model trained entirely on human speech, with no text data, and built from scratch. Most voice systems convert speech to text first, or rely on voice tags to steer emotion. Dessa works directly on audio, which Block says lets it pick up emotion, sarcasm, laughter and pauses in real time. TechCrunch notes it is not clear whether some kind of tag steers that emotion. Dessa is full-duplex, meaning it can listen and talk at the same time, like the GPT-Live model OpenAI released earlier this year. Unlike typical speech models, it is not paired with a large language model for knowledge retrieval, and Block says its knowledge comes from the conversational data it was trained on. The model was trained on a single NVIDIA GB200 NVL72 rack and runs locally on a MacBook Pro M4 Max. Block CEO Jack Dorsey called it "the most authentic full-duplex speech model ever built." Dessa is not publicly available yet. Block says safety concerns were part of its motivation and is first recruiting AI security researchers through a sign-up form to test the model for unexpected risks.