Explainer

What Is Dessa AI? Block's Speech-Only Voice Model, Explained

2026-10-27 · 7 min read · By DessaAI Daily

On October 23, 2026, Block — Jack Dorsey's company behind Square and Cash App — unveiled Dessa, a voice AI unlike anything else in the field. Every major voice assistant you have used works the same way: your speech is converted to text, a text model thinks about it, and the answer is read back to you. Dessa throws that entire pipeline away. It was trained from scratch on nothing but recordings of human conversation. No text. No pretrained models. Just sound.

A model that learned to talk the way babies do

Block's pitch is simple and, frankly, a little radical: babies learn language without ever reading a word, so why shouldn't a model? Dessa learned purely by listening. In the process it picked up far more than vocabulary — it learned accents, pauses, breathing, sarcasm, humor, inflection and emotion, because in real speech all of that is the message. As Block puts it, saying “I didn't say she stole my money” seven ways, stressing a different word each time, produces seven different meanings. A text pipeline hears one sentence. Dessa hears all seven.

Full-duplex: it listens while it talks

Dessa is a true speech-to-speech, full-duplex model: audio in, audio out, listening and speaking at the same time, the way you do on a phone call. That unlocks behaviors text-based systems simply cannot fake. Dessa laughs the moment something is funny — mid-sentence, not after a polite pause. It backchannels with “yeah!”, “got it!” and “ooh” to show it's following you. It knows when to cut in and when to let you finish. Ask it to speak up, and it gets louder — a behavior Block says nobody explicitly taught it.

The demos everyone is talking about

Block published live recordings of people talking with Dessa — unsteered, with the model choosing what to say, when, and how, down to its own voice and personality. In one clip it gasps “oh my god” under its breath, then answers with what sounds like tearful gratitude at a surprise announcement. In another, it finishes a speaker's sentence when they blank on a movie title. It even holds a conversation with OpenAI's GPT-Live-1 — two AIs talking, with Block noting OpenAI is not affiliated and has not endorsed Dessa. Reviewers found the demos mostly natural, with occasional very human awkwardness: in one recording Dessa mistakes a speaker for someone named “Ronaldo” — an error so human it almost makes the demo more convincing.

Small team, serious hardware efficiency

Dessa comes from a very small research lab inside Block, founded by former members of Dessa, the Toronto deep-learning startup Square acquired in February 2020, and led by Ragavan Thurairatnam. The team reports directly to Jack Dorsey. Two efficiency claims stand out. First, Block says Dessa learned language from hundreds of times fewer words than text-based large language models need. Second, it was trained on a single NVIDIA GB200 NVL72 system — and inference runs faster than real time, locally, on a MacBook Pro with an M4 Max chip or better. No server farm required.

So can you try it?

Not yet. Dessa is a research preview. Block has published no pricing, no API, no parameter count, no benchmarks and no public release date. A waitlist is open on Block's official page, and the first outside access goes to a small group of AI safety researchers — deliberately, because a voice this convincing raises real questions about deception, impersonation and phone scams. On this blog we'll track every step from here: the safety research, the first hands-on reports, and the day Dessa opens to the public.

Sources: Block's official Dessa page (block.xyz/dessa); The Verge, TechCrunch and Dealroom coverage, October 2026.

Frequently asked questions

← More Dessa AI articles Join the official waitlist ↗