Block, the company led by Jack Dorsey, has released Dessa, a new voice model from its research lab that runs locally and produces strikingly human expressions such as laughter, pauses and crying. The company said the model was trained entirely on human speech using a single NVIDIA GPU rack.
Dessa is a full-duplex model, able to speak and listen simultaneously, similar to OpenAI’s GPT-Live. While most speech models rely on a paired large language model for knowledge, Block said Dessa’s knowledge comes from its conversational training data.
In a blog post, Block argued that current voice AI often feels unnatural because it converts speech to text, discarding information text cannot carry, so even the most powerful models cannot hear how users feel. The company described this as a worse and potentially dangerous experience. Unlike models such as ElevenLabs v4 and Fish Audio Drama 3, which use voice tags to steer emotion, Block claims emotion is built into Dessa, though it is unclear whether tags play any role.
Researchers shared demos showing varied voices, cadences and emotions. Dessa can run on consumer hardware such as a MacBook M4 Max, but is not yet publicly available while Block invites safety researchers to test it. An on-device model could help Block, whose Square unit already uses voice AI for ordering, power more experiences at lower cost.