Direct answer: Jack Dorsey’s payments company Block has unveiled Dessa, a voice AI model it says was trained from scratch on nothing but recorded human conversation — no text, no pretrained models — and which can listen and talk at the same time, laughing, pausing and interrupting the way a person does. It is a research preview behind a waitlist, not a product you can buy, but it is one of the most interesting voice AI announcements of the year because it challenges the way almost every voice assistant today is built.
What happened
On 23 October 2026, Block announced Dessa, a speech-to-speech model from a small AI research lab inside the company. The lab was founded by former members of Dessa, the Toronto deep-learning startup, and is led by Ragavan Thurairatnam, reporting directly to Block co-founder and chairman Jack Dorsey. Dorsey announced it himself on X, calling Dessa “the most authentic full-duplex speech model ever built” and writing that talking with it “feels like talking with a person… full of emotion, nuance, pauses, and a natural cadence.” His post was viewed more than 6.7 million times.
The name is a small piece of history in itself: the lab takes its name from the Toronto startup its founders came from, and the model inherits it. Block is best known for Square, the card reader and point-of-sale system used by millions of small businesses, and Cash App, its consumer money app. A voice AI lab inside a payments company may sound odd, but it makes more sense than it first appears: Block’s businesses live or die on conversations with customers, sellers and support callers, and voice is the oldest customer-service channel there is.
Block published a set of recorded demos alongside the announcement, including a conversation between Dessa and OpenAI’s GPT-Live-1 voice system. The Verge, which covered the launch, described the demos as mostly natural but occasionally awkward and uncanny, noting one exchange in which the model mistook the speaker for someone named “Ronaldo”. That is worth keeping in mind throughout this article: everything the public currently knows about how Dessa sounds comes from recordings Block chose to publish.
How Dessa actually works
To understand why Dessa is different, it helps to understand how the voice assistant on your phone works today. When you talk to a typical assistant, three separate things happen in a chain. First, a speech-recognition system turns your words into text. Then a language model reads that text and writes a text reply. Then a speech-synthesis system reads that reply aloud. Each step is good at its own job, but each step also throws information away. The text in the middle has no record of whether you sounded happy, sarcastic, tired or angry. It does not know you laughed halfway through a sentence, or that you paused because you were thinking, or that you were interrupted. By the time the answer comes back as speech, the emotional content of the original conversation has been flattened into plain words.
Dessa removes the middle step. Audio goes in and audio comes out, with no transcription layer in between. That single change is the whole bet. Because the model was trained on raw conversation audio rather than text, Block says it learned things that text-trained systems never see: when to laugh, when to lower its volume, when to say “yeah” or “got it” while the other person is still talking, and when to wait. This is what “full-duplex” means in practice. A normal phone call is full-duplex: both people can speak and listen at the same time, interrupt each other, and talk over each other. Most AI voice systems are closer to a walkie-talkie: you speak, then it speaks, and the two never overlap. Dessa is designed to behave like a phone call.
Block’s second striking claim is about how Dessa was trained. The company says it trained the model from scratch on nothing but human conversation, with no text data and no pretrained model underneath. In its own words, “We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models. It learned to hold a real-time conversation simply by listening.” The company also says Dessa learned language from “hundreds of times fewer words” than large language models typically see, while conceding that language models also have to learn things like programming languages, which a speech-only model does not.
The third claim is about efficiency. Block says it trained Dessa on a single NVIDIA GB200 NVL72 system — one rack of Nvidia’s most powerful AI hardware. That is still an enormous amount of computing power, and it is not something a small business could ever replicate. But it is far smaller than the warehouse-scale clusters used to train frontier text models, which can run to tens of thousands of chips. Block’s argument is that training on speech turns out to be far more data-efficient than training on text, because a single second of speech carries far more information than the handful of words it contains: tone, pace, emphasis, breath, emotion and intent are all packed into the sound itself.
Finally, Block says Dessa runs locally on a MacBook Pro with an M4 Max chip or better, with inference faster than real time. If that holds up, it matters for two reasons. The first is privacy: a model that runs on your own machine does not need to send your voice to a cloud server. The second is cost and latency: no round trip to a data centre means no waiting and no per-minute API bill. Both of those claims, however, are Block’s own, and neither has been independently verified yet.
What Block has not shown
Honest coverage of Dessa has to be clear about what is missing, because the gaps are large. Block has not published a parameter count, so nobody outside the company knows how big the model is. It has not published a dataset description, so nobody knows whose conversations it was trained on, how consent was handled, or how representative those conversations are. It has not published benchmark results, latency measurements or a technical paper, so there is no independent way to compare Dessa against rival voice systems on accuracy, speed or safety. There is no API, no pricing, no product, and no open weights. The only public evidence is the demos Block selected.
That is not unusual for a research preview — companies routinely announce research before they have a product — but it does mean every claim in this article that starts with “Block says” should be read exactly that way. The demos are also a limited form of evidence: a handful of polished conversations cannot tell you how a model behaves on a bad phone line, with a strong accent, in a noisy shop, or when a caller is upset. The Verge’s observation that one demo veered into awkwardness is a useful reminder that even the chosen recordings are not flawless.
There is also a safety question that Block itself has raised. The company says it is giving a small group of AI safety researchers early access specifically to find risks it has not anticipated, saying: “A model that learned to talk from sound alone is new territory, and we want outside eyes on it.” That is the right instinct, because a voice model that can mimic human emotion convincingly raises obvious concerns: impersonation, phone scams, social engineering and manipulation all get easier when a machine can sound genuinely warm, hesitant or concerned. For small businesses, which are frequent targets of exactly those scams, this is not an abstract worry. A voice model that can be pointed at a supplier or a customer and trusted because it sounds human is a tool that will need strong guardrails before it is anywhere near a business phone line.
How Dessa compares with today’s voice AI
Dessa arrives into a voice AI market that has moved fast in the last two years. OpenAI’s voice mode, Google’s Gemini Live and a string of startups all offer conversational voice, and several — including Sesame, which made waves with its own highly natural voice demos — have already shown that AI voices can be remarkably lifelike. What sets Dessa apart, on Block’s account, is not that it sounds human — several systems now do — but how it got there: from sound alone, with no text and no pretrained language model, on a single hardware rack, running locally on a laptop.
If those claims survive independent scrutiny, the implications go beyond one model. They would suggest that speech is a far richer training signal than the AI industry has assumed, and that smaller, cheaper, more private voice models are possible than the current cloud-based giants imply. That would be good news for anyone who wants voice AI without sending every conversation to a third party’s servers. If the claims do not survive scrutiny, Dessa still matters as a demonstration of how far speech-only training can be pushed, and as a sign that Block intends to be a serious AI research player rather than just a payments company.
It is also worth noting what Dessa is not. It is not a general-purpose chatbot that happens to talk. Because it was trained on conversation rather than text, it has not learned to write code, solve maths problems or browse the web the way a text language model has. Its strength is conversation itself: the rhythm, emotion and back-and-forth of talking. That specialisation is both its appeal and its limit. A business would not replace a general AI assistant with Dessa; it would use a Dessa-like model for the specific job of talking to people, while other systems handle the knowledge and task work behind the scenes.
What it means for UK small businesses
For most UK small businesses, Dessa changes nothing today. You cannot buy it, plug it into your phone system, or add it to your website. But the direction it points in is directly relevant to how small firms will handle customer conversations in the next few years, and there are practical things worth thinking about now.
The first is phone answering and appointment booking. Missed calls are lost revenue for tradespeople, salons, clinics, restaurants and any business that takes bookings by phone. Today’s AI phone answering tools already exist, but many still sound robotic, struggle with interruptions, and fail when a caller goes off script. A full-duplex model that handles interruptions naturally and hears frustration in a caller’s voice could make automated answering genuinely usable for small firms, rather than a compromise customers tolerate. If Dessa-like models become available as an API, the cost of a virtual receptionist that sounds human could fall sharply.
The second is customer support. Small businesses increasingly handle support over voice notes and phone calls as well as text, and a model that understands tone could triage better: spotting an angry customer and escalating them to a human immediately, rather than making them repeat themselves to a machine. It could also help with accessibility, giving customers who prefer to talk rather than type a smoother way to get help.
The third is privacy and data protection, which matters more in the UK than almost anywhere else. Under UK GDPR, voice recordings are personal data, and sending customer calls to a cloud AI provider creates obligations around consent, retention and processing agreements. A model that runs locally, on a device the business owns, would sidestep much of that complexity. That is one reason Block’s claim about running on a MacBook is worth watching: local voice AI is a much easier compliance story than cloud voice AI, even if the hardware cost is higher up front.
The fourth is fraud. This cuts the other way. As voice AI gets more convincing, small businesses will face more realistic impersonation calls — fake suppliers changing bank details, fake customers placing fraudulent orders, fake bosses instructing payments. The same technology that could answer your phones could also be used to deceive your staff. The practical response is boring but essential: agree a verification habit now, such as calling back on a known number before acting on any payment instruction received by phone, and make sure every member of staff knows it. Technology will not solve this; process will.
None of this requires action this week. The sensible posture for a small business owner is to keep Dessa on a watch list, not a shopping list. When Block announces an API, pricing or a product built on Dessa, that is the moment to evaluate it — ideally by testing it on real calls, with real accents and real background noise, before letting it anywhere near a customer.
Key facts
| Fact | Detail |
|---|---|
| Announced | 23 October 2026 |
| Made by | A small AI research lab inside Block (Square and Cash App), led by Ragavan Thurairatnam, reporting to Jack Dorsey |
| What it is | A full-duplex speech-to-speech model: audio in, audio out, no transcription step |
| Trained on | Recorded human conversation only — no text, no pretrained models, per Block |
| Training hardware | One NVIDIA GB200 NVL72 system, per Block |
| Runs on | A MacBook Pro with an M4 Max chip or better, faster than real time, per Block |
| Access today | Research preview behind a public waitlist; early access for a small group of AI safety researchers |
| Not yet published | Parameter count, dataset details, benchmarks, API, pricing, weights, technical paper |
Timeline
Block’s announcement did not come out of nowhere, but the public timeline is short. The lab behind Dessa was built by people who had previously worked on generative speech at the Toronto startup of the same name, and Block says the team has been working on generative speech models for more than eight years. On 23 October 2026, Dorsey announced the model publicly on X, Block published its Dessa page and demo recordings, and coverage followed the same day from The Verge, The Indian Express and others. Within a day, the announcement had been viewed millions of times and picked up by technology press around the world. The next milestones to watch are the first independent evaluations from the safety researchers Block has given early access to, and any announcement of an API, product or open release.
What to watch
There are four things worth watching in the coming months. The first is independent evaluation: what do the safety researchers Block invited in actually find, and does Block publish their findings? The second is a technical paper or benchmark release, which would let the wider research community judge whether speech-only training really is as efficient as Block claims. The third is product news: does Dessa become part of Square’s seller tools, Cash App, or a developer API? A voice model inside Square’s point-of-sale ecosystem would be the most direct route to small businesses. The fourth is the competitive response: OpenAI, Google and the voice startups will not stand still, and the most interesting question is whether they adopt speech-only training themselves or argue that text-trained models with better voice layers are enough.
For now, Dessa is best understood as a research result with a big claim and a small evidence base. The claim — that a machine can learn to converse from listening alone, the way a child does — is genuinely exciting. The evidence — a handful of company demos — is genuinely thin. Both things can be true at once, and holding both in mind is the right way to follow this story.
FAQs
What is Block’s Dessa?
Dessa is a full-duplex speech-to-speech AI model from a small research lab inside Block, the company behind Square and Cash App. It takes audio in and gives audio out, listening and speaking at the same time, and Block says it was trained from scratch on nothing but recorded human conversation, with no text and no pretrained models.
How is Dessa different from Siri, Alexa or ChatGPT voice mode?
Most voice assistants work in three steps: turn your speech into text, send the text to a language model, then turn the answer back into speech. Dessa skips the text step entirely. Because it works in sound from start to finish, Block says it keeps the emotion, timing, laughter and pauses that text transcription throws away.
Can I try Dessa today?
Not yet, unless you are in the small group of AI safety researchers Block is giving early access to. For everyone else, Dessa is a research preview behind a public waitlist. Block has not announced a release date, pricing, an API or open weights.
Does Dessa run on an ordinary computer?
Block says inference runs faster than real time on a MacBook Pro with an M4 Max chip or better. That is a high-end laptop, not an everyday office machine, and it is Block’s own claim rather than an independently tested result.
What should UK small businesses watch for?
Voice AI that understands tone could eventually change phone answering, appointment booking and customer support. But Dessa is not a product yet. The practical move now is to watch for an API or product announcement, and to make sure any voice tool you do adopt is tested on real customer calls before it touches your customers.
Sources and further reading
Block’s Dessa announcement and demo page, Jack Dorsey’s announcement on X, and coverage from The Verge, The Indian Express and Dealroom informed this article. For background on how voice AI is already reaching small firms, see our guides on AI agents for small businesses in the UK and automating lead follow-up with AI agents. For a related story on voice agents that run locally, see our coverage of Nvidia’s Nemotron Bolt voice model.
