The demo that feels like a phone call
Block built its Dessa voice AI on a complaint that is easy to picture: talk to a modern voice assistant and the seams show within seconds. It waits, politely, for you to finish. Then it answers. The exchange has the rhythm of a walkie-talkie: over, then over, then over. Useful. Civilised. And about as natural as a fax.
Now listen to Dessa. In the clips Jack Dorsey’s Block released this week, the model laughs in the middle of a sentence, the way people do, instead of waiting for its turn. Ask it to speak up and it gets louder. A person blanks on a movie title and it finishes the sentence for them. Tell it about a surprise trip to Paris and it gasps, quietly, and answers with what sounds like teary gratitude.
And then, in one of the flagship demos, it greets a speaker by mistaking him for somebody named “Ronaldo.”
That is the whole story of Dessa in miniature: a machine that talks like an awkward new friend, fluent and warm and occasionally wrong in a way that is unnervingly human. What Block cannot show you, in any of its hand-picked clips, is why it works — or even whether the reason it works is the one Block is selling.
What Block says Dessa voice AI can do
Dessa, announced on October 23, 2026 by Block’s in-house AI research lab, is a speech-to-speech voice model. Audio goes in and audio comes out, with what the company says is no detour through written words along the way. It is full-duplex, so it listens while it speaks, and it has learned, Block says, when to cut in and when to backchannel — “yeah!”, “got it!”, “ooh” — the way people signal they’re following along.
The headline bet is the training. Block says Dessa was built “from scratch on nothing but human conversation.” No books, no web pages, no text of any kind. Just people talking. Dessa lead Ragavan Thurairatnam has described it as the first AI model that can hold a realistic conversation after training entirely on speech, and said: “We made the first AI model you can have a realistic conversation with that was trained entirely on speech!” (reported by The Verge and Unite.AI).
It is a deliberate, almost philosophical choice. Block’s analogy is a toddler: “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.”
The company is careful, to its credit, to describe how hard. A spoken word occupies thousands of times more bytes than the same word in text. Say the same sentence twice, exactly the same way, and the data still comes out different — speech never repeats, where typed text is identical no matter who types it. And yet, Block says, Dessa learned language from hundreds of times fewer words than the big language models, on a single NVIDIA GB200 NVL72 system. “(In fairness,” the page adds, with rare candour for a product announcement, “LLMs also have to learn C++ and whatever the hell Angular is.”)
The impressive parts are real engineering
Strip away the mystique and there is genuinely novel craft here. Full-duplex voice — the phone-call trick of speaking and listening at once — is what makes the demos feel alive, and it is the part nobody should dismiss. Block says inference runs faster than real time, so replies arrive without the awkward pause, and the model runs locally on a MacBook Pro M4 Max or better. It hears and speaks emotion, sarcasm, laughter, a frustrated sigh, a voice cracking with sadness — the prosodic layer that, in Block’s description, most voice AI discards when it “flattens what we say into text.”
This is the part of the story that needs no causal theory. A model that reacts to an interruption mid-sentence is an engineering achievement regardless of how it was trained. The demo of Dessa playing a voice guessing game with a person — turn-taking under real timing pressure, which a scripted question-and-answer session does not test — is a meaningful showcase, and Block deserves credit for showing it.
But every impression comes from clips Block chose
Here is the itch the demos cannot scratch. Dessa talks fluently, shows what looks like feeling, and gets basic facts wrong — and every one of those impressions arrives in clips Block itself selected. The company is upfront about one thing: “Nobody steers it. Dessa picks what to say, when to say it and how, down to its own voice and personality. Subtitles were added afterward.” That is a claim about the demos’ integrity, not a substitute for independent testing.
The causal question hangs over everything. Is the human-ness of the conversation what falls out of training without text — or is that the story being told while the evidence stays inside a few clips? The reported materials do not address whether a text-trained model would make the same mistakes on matched prompts. The Ronaldo moment proves nothing specific about speech-only training: text-trained frontier models also hallucinate names and veer into tangents, as the Nolan-films clip’s detour into Bryan Singer’s Superman films reminds us. And the genuinely novel engineering — full-duplex timing, interruptions, laughter handled live — deserves the headline far more than the mistakes do.
None of this makes Dessa a trick. It makes it an unanswered experiment being presented with the confidence of a concluded one. Block lists the right open questions on its own page — how efficiently such models learn language and facts compared with text, how efficiently they learn new languages, what in-context learning they can do. The demos are designed to impress; what they cannot do is answer the questions they leave hanging.
A tiny lab, a familiar name, a big safety bet
The people behind this are worth noting. The lab is what Block describes as a very small AI research group founded by former members of Dessa, the Toronto deep-learning startup whose name it now wears as a product — a company Block acquired in 2020. The team says it has built deep learning models since 2012 and generative speech models for over eight years, and it reports directly to Dorsey. Lab lead Ragavan Thurairatnam called the release “a GPT3 moment for speech” — the comparison being, presumably, to the release that made many people take text-generating AI seriously.
Block gives three reasons for building Dessa: “curiosity (aka fun), better understanding between humans and AI, and safety.” The safety argument is the most interesting and the least developed. The page argues that today’s AI “keeps gaining power and independence, and it’s still nearly deaf to the emotion in our voices,” and that today’s voice AI “flattens what we say into text and can’t keep up with human timing.” The counter-reading writes itself: a model that sounds convincingly human, reacts to emotion, and handles interruptions raises new, voice-specific questions about impersonation and trust. Dessa is not broadly available. Early access goes to a small group of AI safety researchers tasked with studying the model’s implications and risks, and a public waitlist is open behind them.
The honest verdict
The demos prove Dessa can sound human. They do not show why it does — and the causal story, the speech-only training, remains Block’s claim, not established fact. What is established is the craft: a genuinely fluent, emotionally textured voice model — full-duplex by Block’s account, listening while it speaks — from a tiny lab with a long history in the field — Block says it learned language from hundreds of times fewer words than large language models — and opened first to a small group of AI safety researchers.
That is a substantial story without the mystique. It just isn’t the story Block chose to lead with. The awkward new friend is delightful. The unanswered question is what makes it worth watching — and it will only get answered when someone produces evidence that can test the causal claim.
