Jack Dorsey’s Block has launched Dessa, a new voice model that runs locally on a single NVIDIA GPU and mimics human expressions like laughs, pauses, and cries. The system is trained exclusively on human speech data.
How it works
Dessa operates as a full-duplex model, allowing it to talk and listen simultaneously. This mirrors the approach OpenAI took with its GPT-Live model earlier this year. While most speech systems rely on a separate large language model to gather knowledge, Block states that Dessa draws its information directly from the conversational data it processes.
Current models often sound unnatural because text drives the speech generation pipeline. Services like ElevenLabs v4 and Fish Audio Drama 3 use voice tags to dictate emotion. Block claims Dessa embeds emotion within the model itself rather than relying on external steering tags. It is unclear whether the system uses hidden tags to guide these reactions.
“Interacting with current voice AI can just plain suck. It turns your voice into text and throws away an incredible amount of information text alone can’t carry. The most powerful AIs in the world can’t hear how you feel. We think that’s ridiculous, leads to a worse experience, and is potentially dangerous,” the company stated in a blog post.
Real-world examples
Researchers posted several conversations showing the agent using different voices, cadences, and emotions. In one instance, the bot expressed tiredness after work and its tone reflected that fatigue. Another clip shows the system sounding emotional and nearly breaking down while discussing how comfortable the human operator was making it. A third example features a Block team member playing a game with the voice assistant.
Availability and safety
The model runs locally on consumer-grade hardware such as a MacBook M4 Max or better. Public access is not available yet. Block has issued a call for safety researchers to test the system, likely to identify cases where a voice agent could deceive users.
What it means
Block operates as a finance and small business solutions provider. Square currently uses voice AI for ordering tasks. An on-device model capable of running locally could expand business and consumer applications while reducing costs.
