Block, the payments company run by Jack Dorsey, has unveiled Dessa, a voice AI model it says was trained from scratch on recordings of people talking, with no written text at all. Block says the model listens and talks at the same time and runs on a MacBook Pro, but it has not given a release date.
“We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models,” the team writes on Block’s Dessa page. Block announced the model on X on Friday 23 October, and early access will go first to “a small group of AI safety researchers”.
Block’s claim is a model that learned speech the way babies do
Most voice assistants are built on text. Speech is turned into words, a language model works on the words, and its answer is turned back into sound. Newer speech-to-speech systems, such as ElevenLabs’ Eleven Realtime and SpaceXSI’s Grok Voice 3.0, handle audio more directly. Block’s distinction is what Dessa learned from.
Block says Dessa skips text entirely. “Dessa is a true speech-to-speech model, and everything it knows came from human speech,” the page says. “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 “world’s first” is Block’s claim, and its own wording hedges it.
The team explains the idea simply: “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.”
It listens while it talks and runs on a laptop
Dessa is “full-duplex”, Block says, meaning it can listen and talk at once rather than taking turns. That lets it make small listening noises while someone else speaks, as people do, and Block says inference “runs faster than real time”.
Block says the model “runs locally on a MacBook Pro M4 Max or better”, so a conversation can stay on the device rather than going to a data centre. It also says Dessa learned from “hundreds of times fewer words” than text-based models, and that the team “figured out how to train Dessa on only one NVIDIA GB200 NVL72”, a single rack of Nvidia’s data-centre chips. Block has not given a figure for the comparison, a model size or the amount of audio used.
The demos are Block’s own picks
Block’s page has a handful of recorded conversations. The Verge found them mostly natural but “occasionally a little awkward and uncanny”: in one, the model mistakes the speaker for someone called “Ronaldo”. In another, a person tells Dessa about a surprise trip to Paris and it replies with what Block describes as “teary gratitude”.
One clip has Dessa talking to OpenAI’s GPT-Live-1 voice model, a version of which TSN covered when OpenAI rolled out GPT-Live-1.1 in ChatGPT Voice. The Verge says Dessa “comes across as more natural”. Block notes that OpenAI’s model is “not affiliated with Block and has not reviewed or endorsed Dessa”.
Block frames emotion as a safety problem
Block gives three reasons for the project: curiosity, better understanding between people and AI, and safety. Text-based voice AI, the team argues, “flattens what we say into text and can’t keep up with human timing”. “Imagine being in serious pain while a voice AI has no idea and just keeps going,” it writes. “That’s the safety problem. Dessa is a step toward fixing it.”
The lab behind it is small. Block says it was founded by former members of Dessa, the Toronto deep learning start-up that Square, as Block was then called, bought in 2020, and that it reports directly to Dorsey, who shared the launch on X. The page is signed by Ragavan Thurairatnam, the Dessa lead.
A striking idea that still needs outside numbers
If Block’s description holds, Dessa is a notable result. Training a conversational model on audio alone, on a single rack and small enough for a laptop, runs against the usual assumption that voice AI must be built on a large text model. And the safety argument is a fair one: a system that only sees a transcript misses the tone that tells you someone is upset.
But right now everything rests on Block’s word and Block’s chosen clips. There are no benchmarks, no model size, no data figures and no number behind “hundreds of times fewer words”. The “world’s first” claim comes with Block’s own “as far as we know”. And with access limited to a few safety researchers and no release date, nobody outside the company can yet test whether Dessa follows instructions, gets facts right or copes with noisy rooms and accents.
The useful next steps are clear: the further recordings Block says it will share on X, what the safety researchers report, and whether Block publishes technical detail that lets others compare Dessa with the speech-to-speech models already on the market.
Sources
- Block, “Dessa”
- Block on X, 23 October 2026
- Ragavan Thurairatnam on X, 23 October 2026
- Jack Dorsey on X, 23 October 2026
- The Verge, “Jack Dorsey’s Block made an AI model that talks to you like an awkward new friend”, 23 October 2026
- Block (Square), press release on the acquisition of Dessa, February 2020
Related stories
- ElevenLabs Launches Eleven Realtime Speech-to-Speech Models for Voice Agents
- SpaceXSI Releases Grok Voice 3.0, Leading Artificial Analysis’s Speech-to-Speech Index
- OpenAI Rolls Out GPT-Live-1.1 and New Voices in ChatGPT Voice
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
