Jack Dorsey’s payments company Block has unveiled Dessa, a voice model it says was trained entirely on human conversation, with no text and no pretrained language model behind it. Announced on 23 October, Dessa is a speech to speech system: audio goes in, audio comes out, and nothing in between is converted into written words.
That design is the whole point, according to the company. Most voice assistants turn speech into text, run it through a language model, and read an answer back aloud, a round trip Block says throws away timing, tone and emotion. Dessa, by contrast, listens and speaks at the same time, so it can interrupt, pause, laugh mid sentence, and react when a tone changes. Dorsey said talking with it feels like talking with a person, full of emotion and natural cadence. These are company recordings, not independent tests, and that distinction matters for every claim so far.
Why skipping text matters
Block says Dessa was trained from scratch on nothing but recorded human conversation, using a single Nvidia GB200 NVL72 system, and that it can run locally on a MacBook Pro with an M4 Max chip or better. If that holds up outside the lab, it changes the economics of voice software: a capable model could run on a device a business already owns, with no cloud bill for every conversation.
The safety questions are familiar from recent AI training debates, where labs have had to balance capability against control. Block has kept Dessa as a research preview, with no public access, no disclosed model size, no benchmark results and no release date. Early access is going first to a small group of AI safety researchers, asked to probe risks including cases where a voice agent could trick the person on the other end of a call.
Awe, and a warning
Reaction split in two directions at once. Some listeners said they had never heard anything artificial sound so human, while others warned that the same quality could make phone scams harder to detect. Personal assistants are moving from phones into cars and homes, so a model that understands tone, not just words, would arrive where trust is already assumed.
For Block, the commercial logic is easy to see. The company owns Square and Cash App, and Square uses voice technology to handle ordering, so a local, expressive voice model could power more business and consumer experiences at lower cost. Yet the research preview strategy suggests the company knows the harder test is not whether Dessa sounds human, but whether people can still tell they are not talking to a human. Until independent researchers have tested it, Dessa is a striking claim about where voice technology could go, not proof that it has arrived.
