Jack Dorsey's Block built Dessa, a voice AI that learned to talk from speech alone
It listens and talks at once and runs on a MacBook, but safety researchers get to try it first.
On October 23, 2026, Block unveiled Dessa, a voice AI model that its small research lab trained from scratch on recordings of people talking. No text went in, and no pretrained model either. As far as Block knows, it's the first AI you can hold a realistic live conversation with that learned only from human speech.
It's a research preview. Block's page has a waitlist, and a small group of AI safety researchers gets in first. The page lists no weights, no API and no price.
The lab's lead, Ragavan Thurairatnam, announced it on X a little after noon Eastern, and he didn't undersell it.
We made the first AI model you can have a realistic conversation with that was trained entirely on speech! Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs. This is a GPT3 moment for speech.
What Block says it can do
Dessa is full-duplex, so it listens while it talks. From real conversations it picked up when to cut in and when to backchannel, the small "yeah" and "got it" noises that show a speaker you're following. Replies come faster than real time, Block says, and it runs locally on a MacBook Pro M4 Max or better.
The pitch is about what text throws away. Most voice AI turns your speech into text first, and Block says that strips out the feeling: sarcasm, laughter, a frustrated sigh. Dessa hears and makes those sounds directly. In one of Block's clips it laughs mid-sentence. In another, a person blanks on a movie title and Dessa finishes the sentence for them.
Block says Dessa learned language from hundreds of times fewer words than a text model needs. The whole thing was trained on one NVIDIA GB200 NVL72. (Block's page jokes that text models also have to learn C++ and Angular.)
How it differs from earlier voice models
Talking and listening at once isn't new. The Moshi paper on arXiv called its system the first real-time full-duplex spoken large language model. Moshi, though, predicts text tokens as a prefix to its audio, an "Inner Monologue" its authors say improves the quality of its speech. By Block's account, Dessa has no text step at all.
Thurairatnam's later posts push further. One says Dessa developed world knowledge and learned to reason from audio alone, with a clip of an intern playing a guessing game with it. Block's own page treats how well speech-only models learn facts and new languages as open questions for researchers to dig into.
Who gets to use it
Nobody outside Block can talk to Dessa yet, apart from the people in its recordings. The first in will be AI safety researchers, and Block is asking them to find the risks it hasn't anticipated.
Safety is also part of Block's case for building it. The page asks you to picture someone in serious pain talking to a voice AI that has no idea and just carries on.
AI keeps gaining power and independence, and it’s still nearly deaf to the emotion in our voices.
Block hasn't published a test of whether Dessa does catch distress, so for now that claim rests on demo clips, such as one labeled "Life advice". I'd want numbers before believing a voice model can tell when someone's hurting.
The laptop line matters for who could run it. A model on your own machine could, in principle, hold a conversation that never leaves it. Block's page doesn't promise that, though, and it doesn't say whether a public version would run locally.
One clip has Dessa chatting with ChatGPT Live, with a note that the developer of GPT-Live-1 hasn't reviewed or endorsed Dessa. Block says more recordings will be shared on X.