Jack Dorsey’s Block has released Dessa, a voice model from the company’s research lab that runs locally and sounds strangely human, with expressions like laughs, pauses, and cries. Block says Dessa is trained entirely on human speech and can run on a single NVIDIA GPU. It is a full-duplex model, meaning it can talk and listen at the same time, similar to the GPT-Live model OpenAI released earlier this year. Most speech models are paired with a large language model for knowledge retrieval, but Block says Dessa‘s knowledge comes from the conversational data it has.

Block argues that current speech models sound unnatural and annoying because they have text in the pipeline that drives speech generation. Some models, such as ElevenLabs v4 and Fish Audio Drama 3, use voice tags to drive emotion; Block claims emotion is built into Dessa, though it is not clear if tags are still steering it. In a blog post, the company said that turning voice into text throws away an incredible amount of information text alone cannot carry, and called it potentially dangerous that the most powerful AIs cannot hear how a person feels.

Researchers posted example conversations where Dessa had different voices, cadences, and emotions. In one, the bot said it was tired from work and sounded tired; in another, it sounded emotional and almost broke down while talking about someone making it comfortable; a third shows a team member playing a game with the assistant. Block says the model can run locally on a consumer-grade computer such as a MacBook M4 Max or better.

There is no public access to Dessa yet, and Block has invited safety researchers to test the model, likely to investigate cases where a voice agent could trick the person on the other end. Block is primarily a finance company selling solutions to small businesses. Square already uses voice AI for ordering, and an on-device voice model could support more business and consumer experiences at lower cost.