Jack Dorsey's Block announced the release of Dessa, a voice model built in its research lab that operates on a single NVIDIA GPU and produces human-like expressions such as laughs, pauses and cries.

Dessa works as a full-duplex system, meaning it can speak and listen simultaneously, similar to OpenAI's GPT-Live model released earlier this year. Unlike most speech systems that rely on a large language model for knowledge, Block says Dessa draws its knowledge from the conversational data it was trained on.

Block noted that existing models often sound artificial because they insert text into the pipeline to generate speech. Many solutions, like ElevenLabs v4 or Fish Audio Drama 3, use voice tags to add emotion. The company claims Dessa embeds emotion directly, though it is unclear whether tags still influence it.

"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 firm wrote in a blog post.

Researchers shared several demo conversations showing Dessa adopting different voices, cadences and moods. In one exchange the bot reported being tired from work and sounded genuinely weary.

Another snippet features the assistant speaking about being comforted, sounding emotional and nearly breaking down.

A further example shows a Block employee playing a game with the voice assistant.

Block says the model can run on consumer-grade hardware such as a MacBook M4 Max or better, but it is not currently publicly accessible. The company has invited safety researchers to evaluate the model, likely to explore scenarios where a voice agent might deceive users.

Block, known for its financial services and solutions for small businesses, already uses voice AI in Square for order handling. Owning an on-device voice model could enable more business and consumer applications at lower cost.