Block’s new voice model sounds eerily human

Block has unveiled Dessa, a voice model from its research lab that is meant to run locally and speak with the messy little cues people use all the time: laughs, pauses, even cries. The company says it was trained entirely on human speech and a single NVIDIA GPU. That is the headline trick here. Not just voice, but voice without the usual detour through text.

Dessa is a full-duplex model, which means it can talk and listen at the same time. Block compared that approach to OpenAI’s GPT-Live model, released earlier this year. In the company’s telling, Dessa doesn’t need a separate large language model for knowledge retrieval the way many speech systems do. Its knowledge comes from the conversational data it was trained on.

Block is making a direct argument against the way most voice AI works now. It says current systems often sound odd because text sits in the middle of the pipeline and strips out the emotional detail that speech carries. In its blog post, the company went further, saying that flattening voice into text can lead to a worse experience and could even be dangerous. The examples it shared try to prove the point: a voice that sounds tired when talking about being tired, another that sounds close to breaking down in an emotional exchange, and a team member playing a game with the assistant.

The model is also supposed to run on a consumer machine such as a MacBook M4 Max or better. But there’s no public access yet. Block is instead asking safety researchers to poke at it, which makes sense for a system that could be used to fool people on the other end of a call. For a company built around finance and small-business tools, that matters. Square already uses voice AI for ordering, and a local model could make a lot more business and consumer features cheaper to run.