Block says Dessa learned to talk from speech alone, and it runs on a MacBook
Jack Dorsey’s company, in a page posted Friday: a full-duplex speech model trained from scratch on conversation audio, with no text and no pretrained model, on one Nvidia GB200 NVL72. Jay Peters at The Verge, 7:32 p.m. UTC, listened to the clips and found a model that calls a speaker Ronaldo. Early access is a small set of safety researchers. There is no weight release and no score.
Feature2 minEasy
Priya Shah2 min read
- Block
- Dessa
- Speech
- Jack Dorsey
- Local models

Photo: BlockSource page
Block on Friday published Dessa, a speech model it says was trained from scratch on recordings of people talking, with no text and no pretrained model underneath. The page calls it a full-duplex system: it listens and speaks at the same time, and the company says it learned from conversation when to interrupt and when to answer with a short “yeah” or “got it.” Inference, Block says, runs faster than real time, and the model runs locally on a MacBook Pro with an M4 Max or a faster chip. The team says it trained Dessa on a single Nvidia GB200 NVL72. Jack Dorsey, Block’s cofounder, posted the page on Friday. Jay Peters’s story in The Verge went up at 7:32 p.m. UTC.
The claim Block puts in the largest type is a first. “As far as we know, Dessa is the world’s first AI you can have a realistic live conversation with that was trained only from human speech.” That sentence is the company’s. The page argues the data problem is easy to underestimate: the same sentence said twice still produces different audio, one spoken word takes thousands of times more bytes than the typed word, and stress changes the meaning. It says Dessa learned language from hundreds of times fewer words than a text model needs, and adds that text models also have to learn programming languages. There is no parameter count, no training-set size, and no benchmark table on the page.
The clips are Block’s, and one of them gets the name wrong
What Block did publish is a set of live recordings, with subtitles added later. The page says nobody steers the model in those clips. One is a conversation with OpenAI’s GPT-Live-1. Block’s caption says both voices are generated, and that the developer of GPT-Live-1 is not affiliated with Block and has not reviewed Dessa. Peters listened to a clip shared by Ragavan Thurairatnam, the lab’s lead, and wrote that the back-and-forth is mostly natural, except that the model at first takes the speaker for someone named Ronaldo. Other clips on the page include a movie conversation that wanders, and a surprise-trip story in which, as Block describes it, Dessa answers with what sounds like teary gratitude. The examples were chosen and published by the team. No outside evaluation is attached.
Babies learn to talk without ever seeing a word written down. We wanted to see if a model could too. It can. It’s just very hard.
— Block’s Dessa page, on why the lab built the model
The other two reasons on the page are a better account of how people and models understand each other, and safety. Block’s version of the safety case is that voice systems flatten speech into text and miss emotion, and that a system which cannot hear pain will keep going. “That’s the safety problem. Dessa is a step toward fixing it,” the page says. The lab is small, reports to Dorsey, and says it was founded by former members of Dessa, a Toronto deep-learning startup. Thurairatnam is offering early access to a small group of safety researchers and asks for an institutional email. The page does not give a public release date, a weight file, or a price. More recordings, it says, will be posted later.


