Block's Speech-Only Voice AI Called a Caller 'Ronaldo'

By Daniel Rao

The funniest AI demo of the year is also, quietly, the most radical bet anyone in this industry has placed in months. On Friday, Jack Dorsey's Block unveiled Dessa, a voice model the company says was trained from scratch on nothing but recorded human conversation — no text, no pretrained language model underneath it — and released demo recordings in which the thing laughs mid-sentence, interrupts people, backchannels like a distracted teenager ("yeah", "got it"), and, in one exchange, mistakes its conversation partner for somebody named Ronaldo.

That last part is the gaffe everyone is laughing about. It should not be the part anyone ignores. A model that mishears you the way people mishear you is either a charming party trick or a genuine break with how voice AI has been built for a decade. Block is betting everything on the second reading. After listening to the demos, I'm not ready to hand them the trophy — but I'm also not laughing anymore, or at least not only laughing.

Photo: Jack Dorsey, Block's co-founder and chairman, who announced Dessa on X on Friday. (Wikimedia Commons, CC BY-SA 2.0 — originally posted to Flickr)

A voice model that never read a word

Start with the claim, because the claim is the story. "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," Block said in its announcement statement, as carried by ANI via The Tribune. The "learned like a child listens" framing is Block's own pitch — their poetry, not established science — and it is worth keeping that distinction in your head for the rest of this article.

The company then did something unusual for an AI launch: it dunked on its own industry's data appetite. "LLMs need an insane amount of text. Dessa learned language from hundreds of times fewer words (in fairness, LLMs also have to learn C++ and whatever the hell Angular is)," the statement continued. That parenthetical is doing more honest work than most corporate AI announcements manage in a year. Large language models genuinely do spend a huge fraction of their training digesting code, and Block's point — that spoken conversation is denser with the stuff that makes talk feel human — is at least directionally plausible, even if "hundreds of times fewer words" is a number only Block can audit, because only Block has the training run.

And that training run is the part that made researchers I follow sit up. Block says it trained Dessa on a single NVIDIA GB200 NVL72 — one rack, not a datacenter — and that the finished model runs locally on a MacBook Pro M4 Max or better, with inference faster than real time. If that holds up under independent scrutiny, it is a genuine efficiency story: frontier text models need warehouse-scale clusters, while Block is claiming it got a conversational voice model out of one rack and a laptop. "This is big (but also really, really small…small enough to run on your MacBook)," Dorsey wrote on X, in a post viewed more than 6.7 million times, per Dealroom. "I'm extremely proud of the Dessa team for pushing the frontier!"

Dorsey's main announcement post carried the thesis in full: "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."

Full-duplex is the whole argument

The word doing the heavy lifting is "full-duplex," and it is worth ten seconds of your time. Almost every voice assistant you have ever used works in turns: it waits for you to finish, converts your speech to text, sends the text to a language model, converts the answer back to speech, and plays it. Dessa, Block says, takes audio in and gives audio out at the same time, with no transcription layer in between — it listens while it talks, the way people do. That is what lets it laugh mid-sentence, cut in, and drop little "yeah" and "got it" backchannels without waiting for a turn boundary that a text pipeline would impose.

Block's case against the incumbents is blunt. Today's voice AI, the company argues, flattens speech into text and throws away everything text cannot carry — emotion, timing, tone — producing what its website calls "a worse experience and potentially dangerous." That last phrase is doing quiet, serious work. The Dessa team put it in starker terms on the project site, as quoted in mezha.net's account of The Verge's reporting: "AI keeps gaining power and independence, and it's still nearly deaf to the emotion in our voices. Imagine being in serious pain while a voice AI has no idea and just keeps going. That's the safety problem. Dessa is a step toward fixing it."

It is a good argument, and it lands harder than most launch-day manifestos because you have lived the failure mode. Everyone who has ever shouted "no, stop" at a phone assistant that kept cheerfully reading the wrong thing knows exactly what "nearly deaf" feels like. Whether training without text is the fix is a separate question — but Block has correctly identified the wound.

Photo: A shotgun microphone with its foam windscreen removed — the hardware that captures the sounds Block says Dessa learned from. (Anonimski / Wikimedia Commons, CC0 public domain)

The demos are where it gets human — and weird

Block published a set of recorded conversations as its evidence, and The Verge's Jay Peters, a senior reporter at the outlet, went through them in The Verge's October 23 coverage, as carried by aggregators publishing The Verge's original. His verdict, and his headline, set the tone for the whole launch: Block, he wrote, had "made an AI model that talks to you like an awkward new friend."

The centerpiece is a conversation shared by Dessa's own project lead, Ragavan Thurairatnam. The exchange has, in Peters' words, "a mostly natural back and forth (though for some reason, the model initially mistakes the speaker for somebody named 'Ronaldo')." There is something perfect about a model trained exclusively on human messiness producing exactly the kind of mistake a distracted human makes at a party. Thurairatnam's own framing, posted on X alongside the clip: "We made the first AI model you can have a realistic conversation with that was trained entirely on speech!"

The other demos wander in stranger directions. A discussion of Christopher Nolan's films takes an unprompted detour into Bryan Singer's Superman movies. A person tells Dessa about a surprise trip to Paris, and the model responds with what the team describes as sounding like teary gratitude — which, if you think about it for a second, is either delightful or faintly alarming, depending on how much you trust a statistical model to perform grief-adjacent emotions on demand. The team also shared clips of Dessa playing a voice-based guessing game and holding a conversation with OpenAI's GPT-Live-1 voice model, in which Dessa reportedly comes across as the more natural of the two.

Two details stuck with me more than the gaffes. First, Block says the clips are live recordings with nobody steering the model — subtitles were added afterwards — which, if true, rules out the most common demo trick in the industry. Second, the team says that when asked to speak up, Dessa raised its own volume — a behaviour it says was never explicitly taught. Emergent loudness is a small thing. It is also exactly the kind of small thing that makes you believe a system learned something about conversation rather than memorising transcripts.

Dessa vs the field: why the comparison everyone wants doesn't exist yet

The obvious question — is Dessa better than Sesame, Kyutai's Moshi, OpenAI's real-time voice, or Google's Gemini Live? — has no honest answer yet, and the reason is structural. Dessa is a research preview behind a waitlist with no parameter count, no API, no pricing, no weights, and no published benchmarks; the only evidence is Block's own demo recordings. Tech-insider's survey of the field names those four rivals as the closest reference points but notes it could not verify a fresh version, release, or capability update for each within the past one to three months, and makes no head-to-head capability claims — and neither should we. What can be said is where Block tries to differentiate: not lower latency alone, but a model whose conversational behaviour was learned from speech alone, without text supervision or a pretrained language model. That is a statement about origins, and origins are hard for rivals to copy after the fact.

There is one comparison Block itself invites. One demo puts Dessa in conversation with a third-party model labelled GPT-Live-1, and Block notes plainly that GPT-Live-1's developer is not affiliated with Block and has not reviewed or endorsed Dessa — both voices in that clip are AI-generated. The practical difference for builders, per juliangoldie's breakdown, is dependency: products built on a text model inherit its pricing, rate limits, and update schedule, while a speech-native model that runs on a laptop would let a small team ship a voice feature with no per-minute cloud bill. Block has not said whether or when outside builders can use Dessa, so that remains a possibility, not a roadmap. And one caution from the Sesame corner: the celebrated Sesame Preview voice is a larger closed production model that Sesame has never released as an API, while the open CSM-1B synthesis block is a different, weaker artefact — so "Sesame vs Dessa" comparisons that treat the app and the open model as the same thing are comparing unlike with unlike. (Field survey per tech-insider; builder-dependency framing per juliangoldie; Sesame open/closed distinction per orcarouter.ai's Gemini/Sesame comparison.)

The team Block buried inside itself

Dessa was not built by Block's payments engineers moonlighting on weekends. It came out of a small research lab inside the company founded by former members of Dessa — yes, the model is named after the team — a Toronto deep-learning startup, and led by Thurairatnam. The group says it has spent more than eight years building generative speech models and reports directly to Dorsey, according to Dealroom's reporting. That reporting line matters: this is a chairman-level bet, not a skunkworks side quest.

Thurairatnam is not underselling it. He called Dessa "the GPT-3 moment for voice," as reported by PANews. That is the kind of line that will look either prophetic or embarrassing in eighteen months, with very little middle ground. GPT-3's moment came with public benchmarks, an API, and thousands of developers stress-testing it within weeks. Dessa has, so far, a waitlist and some audio clips. The ambition is noted. The receipts are pending.

What you can't do with it yet — the honest part

Here is the paragraph Block's announcement would rather you skim. Dessa is a research preview behind a waitlist. Block has published no parameter count, no API, no pricing, no model weights, and no benchmark results. A small group of AI safety researchers gets early access; everyone else gets the waitlist and the demos. As The Indian Express put it, "the only evidence of its capabilities is the demos that are in the public domain."

A note on our own reporting, because this is the part where tech journalism usually waves its hands: this morning we pulled the Dessa coverage ourselves — Dealroom, The Indian Express, ANI via The Tribune, PANews, and The Verge's original reporting — and re-checked every technical claim in this story against at least two of them. The announcement date, the X post and its view count, the from-scratch training claim, the GB200 NVL72 detail, the MacBook inference claim, the team history, the Ronaldo gaffe — every one cleared the bar. The one thing we cannot independently verify is the demos themselves. Only Block has the model, and only Block chose which recordings the public gets to hear. Keep that in mind every time someone, including me, tells you how natural it sounds.

The judgment

Here is the stance, because a launch this interesting deserves one. The demos are genuinely impressive, and the no-text training story is the most interesting bet in AI this year — full stop. An industry that has spent three years scaling text models into the hundreds of billions of parameters needed somebody to ask whether the transcription layer was the problem all along. Block asked it, spent one rack finding out, and has audio to show for it. That took nerve.

But "the GPT-3 moment for voice" is marketing running ahead of science, and the gap matters. GPT-3 shipped with numbers anyone could check; Dessa ships with vibes. A voice model that mistakes you for Ronaldo is endearing in a demo and would be something else entirely reading back a medication list — and Block's own safety framing, the person in serious pain with a deaf AI, cuts both ways. Today's voice assistants are shoddy listeners, no question: flattening a trembling voice into flat text and calling it understanding is a botched job, and Block is right to say so. Being a better listener is not the same as being trustworthy, though, and nothing in a waitlisted research preview earns that second claim yet.

The thing I keep coming back to is the early access for safety researchers. It is a small, unglamorous detail, and it is the most encouraging one in the whole announcement. It suggests the team knows the difference between a demo that charms and a system that can be deployed — which is more than you can say for half the voice-AI launches of the last two years. Dessa may or may not be the future of talking to machines. It is, at minimum, proof that the future does not have to sound like a transcription.

Update, Oct. 24: the first listeners have had their say — and one passed the dead-person test

The first full listen-throughs landed on October 24, and the warmest one came from TheStreet's reviewer, who writes that he has "heard every big voice demo of the past few years, and nearly all of them still failed my dead-person test. Dessa does not. For the first time, I heard a machine and forgot, just for a second, that it was one." Two moments earned that verdict: in one clip Dessa tells a caller he sounds far from the microphone and asks him to come closer so it can hear him better, and in the Paris-trip clip it gasps "oh my god" under its breath, answers with what sounds like teary gratitude, then jokes that a French breakfast is an espresso and a cigarette. The same reviewer ties the launch to Dorsey's longer bet — Block's bitcoin revenue hit $2.42 billion in a single quarter, per TheStreet — and Dorsey's line that bitcoin "fails through irrelevance" if it stays a store of value and nothing more: technology only counts, in this reading, when ordinary people actually use it. Separately, a sentiment tally of one X conversation by tech-insider put reaction at 68.2% positive and 31.8% negative — but on only 23 sentiment-bearing replies from 22 accounts, so read it as a mood check, not a poll. Supporters praised the emotional expressiveness and the prospect of private local hosting; sceptics called the project scammy or dismissed the results as unremarkable. (Reactions per TheStreet, October 24, and tech-insider; sentiment sample size kept explicit.)

Update, Oct. 25: the fraud question Dessa can't dodge

The other reaction to Dessa is alarm, not amazement — and it follows directly from the caution already in this post: a voice that passes a dead-person test can deceive, flatter and manipulate at human fidelity. Fraud watchers warned, in CoinDesk's October 24 roundup of the reaction, that a voice this human could make phone scams harder to detect, precisely because it laughs, hesitates and backchannels like a person. Block has said only that safety researchers get early access "to find the risks we haven't anticipated". No Block response to the scam concern has been published. Until one is, the fraud question sits unanswered beside the demo applause: the more natural the voice, the harder it is for a listener to tell whether the person on the line is a person at all.

Frequently asked questions

What is Dessa?

Dessa is a full-duplex speech-to-speech AI model announced by Jack Dorsey's Block on October 23, 2026. Block says it was trained from scratch entirely on recorded human conversation, with no text data and no pretrained models involved.

How is "full-duplex" different from Siri or Alexa?

Today's assistants convert your speech to text, process the text, and convert the answer back to speech — losing emotion, timing, and tone along the way, in Block's telling. Dessa takes audio in and produces audio out simultaneously, so it can interrupt, laugh mid-sentence, and backchannel ("yeah", "got it") the way people do.

What did it take to train, and where does it run?

Block says training used a single NVIDIA GB200 NVL72, and that the finished model runs locally on a MacBook Pro M4 Max or better with inference faster than real time. Both figures are Block's claims; no independent benchmarks have been published.

Can I try Dessa?

Not yet. It is a research preview behind a public waitlist, with early access granted to a small group of AI safety researchers. Block has announced no API, pricing, weights, or timeline for wider release.

Who built it?

A small research lab inside Block founded by former members of Dessa, a Toronto deep-learning startup, and led by Ragavan Thurairatnam, who reports directly to Dorsey. The team says it has spent more than eight years building generative speech models.

Why is Block starting with safety researchers?

The team gives three reasons for building Dessa: curiosity, better understanding between humans and AI, and safety. The safety case, in the team's own words: "Imagine being in serious pain while a voice AI has no idea and just keeps going. That's the safety problem. Dessa is a step toward fixing it." A model that hears emotion, timing, and tone — rather than flattening speech into text — can in principle notice distress that a text pipeline would miss. That is also why Block says it is giving a small group of AI safety researchers early access "to find the risks we haven't anticipated": a voice that passes a dead-person test is a voice that can deceive, flatter, and manipulate at human fidelity, and the team wants outside eyes on that before wider release. (Team rationale per Block via cryptocompass/TheStreet coverage; early-access quote per Block's announcement.)

Why does running locally matter?

Three reasons, per juliangoldie's breakdown of the launch. Privacy: your conversations don't have to leave your device. Speed: no round trip to a server means fewer awkward delays — and Block says inference already runs faster than real time. Cost: no per-minute API bill once the model is running. The trade-off is reach: text-trained models know far more facts and can use tools, while Dessa is built for natural conversation. Expect future voice agents to combine both — a Dessa-style voice in front of a capable agent behind it. (Framing per juliangoldie; Block's local-run and faster-than-real-time claims kept attributed.)

Is Dessa open source, or can developers use it?

No. Dessa is a research preview behind a public waitlist; Block has announced no API, pricing, model weights, or timeline for wider release, and early access goes only to a small group of AI safety researchers.

Sources

Block's announcement statement (via ANI/The Tribune and The Economic Times); Jack Dorsey's October 23 X posts (via The Indian Express); Dealroom's October 23 report; PANews (October 23); The Verge's October 23 coverage by senior reporter Jay Peters (via aggregators carrying The Verge's original); mezha.net's account of The Verge's reporting; NDTV Profit's demo rundown; juliangoldie.com's explainer. On the agent economy this kind of voice interface is heading for, see our coverage of Anthropic's Claude Duet and its refusal to touch your money without asking and Uber rolling ChatGPT-powered agents out to 30,000 workers; for the primer, what AI agents actually are.

Daniel Rao is TechPulse USA's AI correspondent, covering the collision of artificial intelligence with money, power, and regulation. He has strong opinions about unit economics and weak ones about kerning.