CYNDRAWhat is Dessa AI?
Dessa is a speech-to-speech AI model from Block, the company behind Square and Cash App. Block announced it on 23 October 2026 with a page of recorded conversations and one sentence from Jack Dorsey calling it "the most authentic full-duplex speech model ever built."
Two things make it different from the voice assistants you already know.
First, Block says Dessa was trained "from scratch on nothing but human conversation." No text. No pretrained language model underneath. Everything it knows, it learned from listening to people talk. Block's page claims it got there with hundreds of times fewer words than a large language model sees, and on a single NVIDIA GB200 NVL72 rack.
Second, it is full-duplex. It listens while it talks. It can say "yeah" and "got it" while you are mid-sentence, cut in when it has something to add, laugh, sigh, and change its tone. Block says inference runs faster than real time and that the model runs locally on a MacBook Pro with an M4 Max chip.
I played the anniversary clip on Block's page twice. The pauses are right. The breathing is right. When the model gets emotional, it sounds like a person getting emotional, not a text-to-speech engine reading a stage direction. That is the part that spread.
Where the name comes from
If "Dessa" sounds familiar, it should. Dessa was a Toronto deep learning startup. In 2019 its team built a synthetic Joe Rogan voice that was convincing enough to make the rounds as a deepfake warning. In February 2020 Square acquired the company.
Block's Dessa page says the lab behind the new model was founded by former members of that startup, that it has built generative speech models for more than eight years, and that it reports directly to Dorsey. The lead is Ragavan Thurairatnam, one of the original Dessa team. So this is not a side project that appeared last quarter. It is a speech team that has been at this since before most of the current voice products existed, inside a payments company that has an obvious reason to care about phone calls.
Why training on speech alone matters
Almost every voice AI you have talked to is a relay race. Your speech gets transcribed into text. The text goes to a language model. The model writes a reply. A text-to-speech engine reads it out. Four handoffs, and the first one throws away most of what makes speech useful: tone, hesitation, sarcasm, the difference between "I'm fine" and "I'm fine."
Block's framing is blunt. Today's voice AI "flattens what we say into text and can't keep up with human timing." Dessa skips the relay. Audio in, audio out, one model. That is why it can react to how you said something, not only to the words.
The same design is also its biggest limitation. A model whose entire knowledge came from recorded conversations does not know what a text model knows. The Verge noticed Dessa wandering off into a tangent about Bryan Singer's Superman films in a demo about Christopher Nolan, and mistaking a speaker for someone called Ronaldo. Hacker News commenters caught it calling Memento "The Mento." One of them summed the thread up as "lots of style, not much substance." That is harsh, but it is also the right question to ask of any demo: what does it do after it sounds good?
Full-duplex is not new. Doing it this well might be.
Dessa is not the first full-duplex voice model. OpenAI shipped GPT-Live-1 inside ChatGPT in July 2026, and it also listens and speaks at the same time. Block knows this. One of the recordings on its page is Dessa talking to GPT-Live-1, and Block is careful to note that OpenAI has no affiliation with Dessa and has not endorsed it.
The difference is in what sits behind the voice. GPT-Live hands harder requests to a text model in the background and feeds the answer back into the conversation. Dessa has nothing behind it. It is the voice and the brain in one network, which is why it sounds so natural and why it sometimes says nonsense with great confidence.
Two caveats worth keeping in mind. Every recording so far was selected and published by Block. Nobody outside the company has run it. And as TechCrunch pointed out, it is not clear whether the emotion in the demos emerges on its own or is being steered by some kind of tag. Those are answerable questions. They have not been answered yet.
Can you use Dessa AI today?
No. If you searched "Dessa AI how to use," "Dessa AI waitlist," or "Dessa AI GitHub," here is the full state of play on 25 October 2026:
- Public access: none. There is no app, no demo you can talk to, no download.
- Waitlist: yes, on Block's Dessa page. The access form asks for an institutional email.
- Early access: a small group of AI safety researchers. Block is asking researchers to get in touch, which TechCrunch read as an effort to find the ways a voice this convincing could be used to deceive people.
- API, pricing, model weights, parameter count, benchmarks: none published.
A Hacker News commenter wrote that Dorsey said the weights will be released. We looked and could not find that statement from Block or Dorsey, so treat it as a rumour until the company says it.
My own guess, and it is only a guess: some kind of developer access in 2027, and a long gap before anything you would put in front of a customer. Block is a payments company. A voice model that can be fooled, or can fool, is a liability for them before it is a product. I would not put Dessa on a roadmap.
What Dessa means if you run voice AI in a business
This is the part I care about, because our work is building AI inside operating companies, and phone calls are where a lot of that work lands: inbound intake, scheduling, collections, order status, follow-up.
The voice was never the hard part. When a voice agent fails in production it is almost never because it sounded robotic. It fails because it could not find the order, booked the wrong slot, did not know when to hand off to a human, or said something it was not allowed to say. Dessa does none of that work. There is no tool calling, no system prompt, no way to connect it to a calendar or a CRM. Commenters who build voice systems for a living listed exactly those gaps, plus English-only support, as the reasons it cannot replace today's pipeline yet. If you want the plumbing side explained, our guide to production-grade AI call assistants covers it.
But the flattening problem is real, and it costs you money. A customer who calls about a late invoice and is angry sounds angry. A tenant who calls about a leak and is scared sounds scared. A pipeline that turns both into neutral text treats them identically, and your escalation rules never fire. A model that hears tone natively could route those calls differently. That is the business case for speech-native models, and it is the one worth watching.
Build so the voice layer is replaceable. This is the practical takeaway. Whatever you deploy this year will use a relay pipeline, or a hosted full-duplex model like GPT-Live. Fine. Put your effort into the parts you will keep: the data the agent can read, the actions it is allowed to take, the hand-off rules, the call logs, the compliance checks. Start read-only and widen permissions as the logs earn trust. Then the voice model becomes a line item you can swap when something better arrives, which it will. If outbound calling is on your list, read how to automate outbound calls without breaking compliance before anything else.
About the FaceTime-with-an-AI idea. The question I actually started with was whether Dessa plus a real-time avatar gets you a video call with an AI that feels like a person. The avatar side exists: HeyGen's LiveAvatar runs two-way video sessions through an API, and the usual setup is to feed it text from a language model, which it then speaks and lip-syncs. Dessa produces audio, not text, so a natural pairing would have to animate the face from the live audio stream. As far as the public documentation shows, nobody has built that, and nobody can until Dessa ships. A believable 20-minute video conversation with an AI that also remembers you and gets work done is a 2027-plus question. The useful work is in the much less glamorous layer underneath: whether the thing on the other end of the call can actually act, not just respond.
What we would do this quarter
If voice AI is on your list for the next six months, here is the order we would work in.
- Pick one call type with volume and a clear outcome. Appointment confirmations, payment reminders, first-line intake. Not "customer service."
- Write down what the agent may read, what it may do, and when it must hand off. That document outlives any model choice.
- Ship it with today's stack, read-only first, and listen to the recordings weekly. The failures will be in the plumbing, not the voice.
- Keep a short list of models to swap in. Dessa goes on the list the day Block publishes an API or weights, and not before.
Dessa is a real step. The speech team at Block showed that a model can learn to hold a conversation from speech alone, and that it can sound more like a person than anything most of us have heard from a machine. That is worth a weekend of attention. It is not yet worth a budget line. The companies that will benefit first are the ones whose voice workflows are already built, logged, and permissioned, so that when a better voice arrives, plugging it in takes a week instead of a quarter. If you would like help getting to that point, we will build the first version with your team in a free live session, on one real call type, inside the tools you already use.
Frequently asked questions
What is Dessa AI?
Dessa is a speech-to-speech conversation model announced by Block on 23 October 2026. Block says it was trained from scratch on recorded human conversation, with no text and no pretrained language model, and that it is full-duplex, so it listens and speaks at the same time.
Who built Dessa?
A small research lab inside Block, led by Ragavan Thurairatnam and reporting to Jack Dorsey. The lab was founded by former members of Dessa, the Toronto deep learning startup Square acquired in February 2020, which is where the name comes from.
Can I use or try Dessa AI today?
Not yet. As of 25 October 2026 there is no public app, API, or download. Block is giving early access to a small group of AI safety researchers, and everyone else can join a waitlist on block.xyz/dessa.
Is Dessa open source? Are the weights on GitHub?
No. Block has not released model weights, a parameter count, pricing, or benchmark results. Comments online claim weights will be released later, but we could not find a statement from Block confirming that.
How is Dessa different from OpenAI's GPT-Live?
Both are full-duplex voice models that listen and talk at once. GPT-Live-1 shipped inside ChatGPT in July 2026 and hands harder requests to a text model in the background. Dessa was trained on speech only, has no text model behind it, and is not publicly available.
Can Dessa run locally?
Block says Dessa runs on a MacBook Pro with an M4 Max chip or better, and its lead has said it can run on consumer GPUs. Nobody outside Block has been able to verify that yet, because the model has not been released.