Image: TechcityAuthority
Most voice assistants are secretly typists. They turn your speech into text, think in text, then read a reply back out loud. Jack Dorsey's Block says it has built one that skips the typing entirely. It is called Dessa, and according to Block, it learned to talk the way babies do, by listening.
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
— jack (@jack) October 23, 2026
Dorsey announced the model on X on October 23, calling it the most authentic full-duplex speech model ever built. Full duplex means it listens and talks at the same time, like a person on a phone call. Block says Dessa was trained from scratch on recordings of human conversation alone. There was no text and no pretrained model underneath.
Put plainly, the samples sound incredible. Dessa laughs mid-sentence when something is funny. It drops in little cues like "yeah" and "ooh" while the other person is still talking. In one clip on Block's page, it gasps "oh my god" at an anniversary surprise.
The first sample below comes from Ragavan Thurairatnam, who leads the Dessa team. As The Verge noticed, the model at first mistakes the speaker for someone named Ronaldo. That slip is oddly reassuring. It sounds like a real, unpolished system, not a staged script.
We made the first AI model you can have a realistic conversation with that was trained entirely on speech!
— Ragavan Thurairatnam (@RagavanThurair) October 23, 2026
Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs.
This is a GPT3 moment for speech.
The second sample is Block's short launch clip. A third shows a Block intern playing a guessing game, which Block says Dessa wins quickly.
Meet Dessa, our speech model trained from scratch on human conversation.
— Block (@blocks) October 23, 2026
Hear it for yourself:
For the love of the game.
— Block (@blocks) October 23, 2026
Watch Cara play a guessing game with Dessa. The model guesses the first one right straight away, and the second doesn’t take long either.
The interesting part is not the voice. It is where the knowledge came from. Every fact Dessa knows, it picked up from raw sound. Block says it used hundreds of times fewer words than text models need, and trained it on a single NVIDIA GB200 NVL72 rack. If those figures hold, speech may be a far richer teacher than anyone assumed, because tone, pauses and stress carry meaning that text strips away.
That is also Block's safety pitch. A text pipeline cannot hear a voice cracking with pain. Dessa, the company argues, can. Fittingly, Block is giving early access first to a small group of AI safety researchers, not the public.
There are big gaps, though. Block has shared no paper, no benchmarks and no release date. Every clip was picked by Block. Thurairatnam calls it a GPT-3 moment for speech and says it runs on consumer GPUs. The Dessa page sets the bar at a MacBook Pro M4 Max or better. Unite.AI also notes a public waitlist. Until outsiders talk to it live, the clips are a promise, not proof.
Still, it is a strange and exciting turn for a payments company best known for Square and Cash App. The team behind it comes from the Toronto startup also named Dessa, and it reports straight to Dorsey. If the model is as good as it sounds, every typist assistant just got a rival that never learned to read.
Comments