Block's Dessa: A Voice AI That Learned From Listening — Waitlist Open

Quick answer

• What it is: Dessa, a full-duplex speech-to-speech AI model announced by Block's in-house research lab on October 23, 2026 — audio in, audio out, with no text anywhere in the pipeline.
• The headline: Block says it was trained from scratch on human conversation audio alone — no text, no pretrained model — learning language from "hundreds of times fewer words" than text models.
• Access right now: a public waitlist at block.xyz/dessa; early access is reserved for a small group of AI safety researchers who apply with an institutional email.
• The hardware surprise: Block says training ran on a single NVIDIA GB200 NVL72 rack, and the model runs faster than real time on a MacBook Pro M4 Max or better.
• What's missing: no public release date, no pricing, no open weights, no API or app — and no independent evaluation of any demo.

Most voice assistants work in three steps: turn your speech into text, feed the text to a language model, then turn the answer back into speech. Block's Dessa skips all three steps. Announced on October 23, 2026, it's a single model that takes in sound and puts out sound — listening and talking at the same time. And according to Block, it learned to do all of it the way a child does: by listening, never by reading.

What Dessa actually is

Block calls Dessa a "full-duplex" speech-to-speech model — audio in, audio out. It listens and speaks simultaneously, so you can interrupt it, laugh over it, or say "mm-hmm" while it's mid-sentence, and it reacts in real time: cutting in, backing off, backchanneling the way people do in ordinary conversation. The company also says Dessa natively hears and speaks emotion, sarcasm, laughter, a frustrated sigh, and a voice cracking with sadness. Inference runs faster than real time, so replies arrive without the awkward pause that usually gives a voice assistant away. Block's own description, which should be read as the company's claim rather than a tested result, is that Dessa is "the most authentic full-duplex speech model ever built" and that talking with it "feels like talking with a person, full of emotion, nuance, pauses and a natural rhythm."

The training story is the real headline

The lab trained Dessa from scratch on nothing but human conversation, according to the announcement. No text training data, no existing language model underneath. "Babies learn to talk without ever seeing a word written down," the company said. "We wanted to see if a model could too. It can. It's just very hard."

The difficulty is worth spelling out because it's easy to underestimate. Speech data never repeats: saying the same sentence twice in exactly the same way produces different data, whereas typed text is identical no matter who types it. Block's page puts it in data terms — one spoken word occupies thousands of times more bytes than its written form — and adds that stress placement changes meaning, with one example sentence taking on a different meaning each time a different word is stressed. The pauses, laughs, overlaps and back-and-forth that make real conversation sound real are all in the audio; the moment speech gets converted to text, that information is gone.

Block says Dessa learned language from "hundreds of times fewer words" than large language models see, and that the team worked out how to train it on a single NVIDIA GB200 NVL72 rack — 72 GPUs, substantial for a single project but small next to the thousands-of-GPU clusters frontier text models train on. That's part of the argument the company is making: speech-focused AI may require a leaner approach to compute. The page also lists open research questions the approach raises — how efficiently such models learn language and facts compared with text, how efficiently they can learn new languages, what in-context learning they can do.

What Block showed — and what the demos give away

Block published clips it describes as live recordings of people talking with Dessa, and the most telling details are the awkward ones. In one conversation the model holds up a mostly natural back-and-forth but initially mistakes the speaker for somebody named "Ronaldo." During a discussion of Christopher Nolan's films, it unexpectedly shifts to Bryan Singer's Superman films. In other clips Dessa laughs in the middle of a sentence rather than after the speaker finishes, gets louder when asked to speak up, finishes a person's sentence when the person blanks on a movie title, plays a voice-based guessing game, and gasps under its breath at a surprise anniversary announcement before answering with what the page describes as teary gratitude. Block says no one steers the model during these conversations: Dessa itself chooses what to say, when to say it, and how to say it, including its own voice and personality, with subtitles added after the fact.

The strangest demo is a conversation between Dessa and OpenAI's GPT-Live-1 voice model. Block notes that both voices are AI-generated, and that GPT-Live-1's developer is not affiliated with Block and has not reviewed or endorsed Dessa. The examples were selected and published by Block, and the model's performance has not been independently evaluated — so treat them as the company's highlight reel, awkward moments included.

The lab's lead, Ragavan Thurairatnam, called the release "the GPT-3 moment for voice" — a big claim for a product nobody outside a handful of researchers can touch yet.

Who built it

Dessa comes from what Block describes as a very small AI research lab inside the company, founded by former members of Dessa — the Toronto deep learning startup whose name the project reuses. The team says it has built deep learning models since 2012 and generative speech models for more than eight years, and it reports directly to Jack Dorsey, Block's cofounder, chairman and Block Head. Thurairatnam previously founded and sold the original Dessa and lectured at the University of Toronto. The reporting line to Dorsey is notable: this is not a side project tucked inside a product org, it's a research effort with the CEO's name on it.

How to join the Dessa waitlist

To sign up:

1. Go to block.xyz/dessa — the only official place to sign up. The page carries the announcement, the audio demos and the waitlist form.
2. Find the waitlist form on the page and decide which track you're applying for: the early-access request or the general public waitlist.
3. Use your institutional email if you're applying as a researcher — from a university, company or lab — with an optional Google Scholar or blog link. Everyone else joins the general list with a normal email.
4. Read Block's privacy notice, then submit.
5. Ignore any message that asks you to pay for access. Block hasn't announced pricing, and popular launches attract fake waitlists and "early access" scams. Only enter your details on Block's own site.

Who gets access first, and why the order is deliberate

Early access goes to a small group of AI safety researchers, and that's not a marketing choice. A voice that convincing raises real questions about impersonation and trust — a model that can laugh, sigh, interrupt and sound genuinely sad is also a model that can deceive. Starting with safety researchers lets Block see how Dessa could be misused before releasing it widely. Researchers should say plainly what they'll study — Block is choosing safety researchers specifically, so a vague application is a weak one.

The team gives three reasons for building Dessa at all: curiosity, better understanding between humans and AI, and safety. On safety, Block argues that AI keeps gaining power while remaining unable to register the emotion in a speaker's voice — giving the example of a person in serious pain while a voice AI has no idea and keeps going — and describes Dessa as a step toward fixing that. Everyone else joins the public waitlist to hear when access opens up; Block hasn't said when that will be, and the page states only that launch links are "coming soon."

What's still unknown: an honest inventory

• Release date — none announced for the public.
• Pricing — not announced. Be wary of anyone asking you to pay for early access.
• Open weights or licence — not announced. A model that runs on a MacBook Pro M4 Max is a natural candidate for a downloadable release, but Block has said nothing.
• Model size — not shared yet.
• API or app — nothing public yet. No developer surface at all.
• Independent evaluation — none exists. Every number and every demo comes from Block.

Two of those unknowns deserve extra caution. The "most authentic full-duplex speech model ever built" line is Block's own marketing until an independent evaluator says otherwise. And the Ronaldo slip and the Superman detour are reminders that a model trained on conversation audio inherits the wandering, mistake-prone texture of conversation — naturalness is not the same as reliability.

What to use instead while you wait

You don't need Dessa to start using voice AI. The gap between "watchable" and "usable" is real, and these are the tools that are usable today:

The honest comparison: Voxtral 3 and ElevenLabs' stack are products with price lists and uptime; Dessa is a research preview with demos. If you're building a voice feature this week, build on what's shippable.

FAQ

Where is the Dessa waitlist?

On Block's official page, block.xyz/dessa — the only official place to sign up.

Who gets Dessa early access?

A small group of AI safety researchers who apply with an institutional email from a university, company or lab, with an optional Google Scholar or blog link — chosen so safety experts can study the model before wider release.

When will Dessa be released to the public?

Block hasn't announced a public release date. The page states only that launch links are "coming soon," with more demos being shared on lead researcher Ragavan Thurairatnam's X account.

Does the Dessa waitlist cost anything?

No pricing has been announced at all — not for access, and not for whatever product follows. Be wary of anyone asking you to pay for early access.

Can I run Dessa on my own MacBook?

Not yet, but Block says the model runs faster than real time on a MacBook Pro M4 Max or better.

Key takeaways

• Block announced Dessa on October 23, 2026: a full-duplex speech-to-speech model its lab says was trained from scratch on human conversation audio alone — no text, no pretrained model.
• Full-duplex means it listens and talks simultaneously, with interruptions, laughter, backchanneling and emotion that text-pipeline voice assistants lose.
• Training ran on a single NVIDIA GB200 NVL72 rack, and the model runs faster than real time on a MacBook Pro M4 Max — a lean hardware story for a frontier-class claim.
• Access is a public waitlist at block.xyz/dessa, with early access reserved for a small group of AI safety researchers applying with institutional emails.
• No release date, pricing, open weights, model size or API has been announced, and nothing has been independently evaluated — treat every claim as Block's own until outsiders verify it.

October 24, 2026 — Dessa is the most interesting voice launch of the month precisely because it questions the pipeline everyone else uses. The demos are charming and the training story is genuinely novel. But the measure that matters — strangers using it on hardware they own — doesn't exist yet.