Jack Dorsey’s Block Builds a Voice AI Trained Only on Speech

Block, the company behind Square and Cash App, has unveiled Dessa, a speech model its research lab says it trained only on recordings of human conversation, according to the Dessa page on Block’s website read on 23 October 2026. “We trained Dessa from scratch on nothing but human conversation,” the page says. “No text, no pretrained models.” It adds that the team trained the model on one NVIDIA GB200 NVL72 rack and that it runs locally on a MacBook Pro M4 Max or better.

Jack Dorsey, Block’s co-founder, announced it on X the same day: “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.” The page offers a waitlist and early access for a small group of AI safety researchers.

Business Pill · ON-DEVICE AI

A short Business Engineer explainer on on-device AI: a model that runs on the device itself instead of in the cloud. It teaches the general idea only and says nothing about Block or Dessa.

What Block Says Dessa Is

Block’s page contrasts Dessa with voice systems that run speech through text, and describes it this way: “Audio in, audio out. Dessa is a true speech-to-speech model, and everything it knows came from human speech.”

“Full-duplex” is the page’s term for a model that does not take turns like a walkie-talkie. In its words: “It’s full-duplex, so it listens and talks at the same time, like you do.” The page says the model learned when to interrupt and when to say things like “yeah!” to show it is following, and that “Inference runs faster than real time, so there’s no awkward wait.”

Source: Block’s Dessa page, block.xyz/dessa, read on 23 October 2026. These are the company’s own descriptions; we did not test the model.

What Was Posted on X

Jack Dorsey announced Dessa on X, writing that voice is the most natural way to interact with intelligence and calling it the most authentic full-duplex speech model ever built; the claim is his own.

How Block Says It Was Trained

The page argues that turning speech into text loses meaning. Today’s voice AI, it says, “turns your voice into text and throws away an incredible amount of information text alone can’t carry.” Dessa, by its account, “learned to hold a real-time conversation simply by listening.”

It also makes a claim about efficiency: “Dessa learned language from hundreds of times fewer words” than large language models need. And it frames the result carefully: “As far as we know, Dessa is the world’s first AI you can have a realistic live conversation with that was trained only from human speech.” The page publishes recorded conversations, including one with GPT-Live-1, which it describes as a third-party AI model, and notes that “The developer of GPT-Live-1 is not affiliated with Block and has not reviewed or endorsed Dessa.”

Who Built It

The team describes itself in one line: “We’re a (very) small AI research lab at Block, founded by former members of Dessa, the Toronto deep learning startup.” It says “We’ve built deep learning models since 2012 and generative speech models for over eight years,” and that “we report directly to Jack Dorsey, Block’s cofounder, Chairman and Block Head.” The page is signed by Ragavan Thurairatnam as Dessa lead and lists six researchers.

It gives three reasons for the work: “curiosity (aka fun), better understanding between humans and AI, and safety.” On safety it says AI keeps gaining power and independence, “and it’s still nearly deaf to the emotion in our voices,” and asks outside researchers to test the model: “We’re giving a small group of AI safety researchers early access to find the risks we haven’t anticipated.”

Connecting the Coverage

In our Voxtral 3 piece, Mistral described a voice agent built as a pipeline in which one model transcribes, a reasoning model thinks and a speech model replies. In our Kimi K3.1 Pro piece, Moonshot listed $2.00 per million input tokens and $10.00 per million output. In our Mubit piece, a start-up offered to train small models on a company’s own agent calls for $5 per 1,000 runs and let it download them.

And in our Universal Cart piece and our bank agents piece, Google put a cart inside Gemini while banks moved to restrict outside AI agents using customers’ logins. What’s new: a payments company is building its own speech model, from audio alone, inside its own lab.

The Structural Read

As we read it, the cost claim is the compute: the page says the team trained Dessa on only one NVIDIA GB200 NVL72, and that it runs locally on a MacBook Pro M4 Max or better.

As we read it, the product question is still open: the page offers a waitlist and early access for safety researchers, and names no price, release date or plan for Square or Cash App.

As we read it, the pitch is the interface: Dorsey’s post calls voice “the most natural way to interact with intelligence”, and the page says Dessa listens and talks at the same time.

Three Implications

FOR VOICE AI BUILDERS The page contrasts Dessa with systems that run speech through text, which it says throw away information text alone cannot carry.

FOR BLOCK The page says the lab reports directly to Jack Dorsey, Block’s co-founder and chairman, and was founded by former members of the Toronto startup Dessa.

FOR SAFETY RESEARCHERS The page says Block is giving a small group of AI safety researchers early access to find risks it has not anticipated.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework says “what changes between paradigms is where intelligence accumulates and who can capture it.”

As we read it, Dessa is Block building its own model for the voice interface instead of renting one. The framework was published in July 2026 and does not discuss Block or Dessa.

What Is Not Established

The Dessa page and Dorsey’s post that we read do not give a price, a release date, a licence, the model’s size or any benchmark scores, and they do not say Dessa will ship in Square, Cash App or other Block products. The page’s privacy notice says it does not cover Square, Cash App, Afterpay or TIDAL. The conversation recordings are Block’s own selection. We read the page on 23 October 2026, did not test the model and did not contact Block. Nothing here is investment advice.

The Bottom Line

Block says its small research lab trained Dessa from scratch on human conversation alone, with no text and no pretrained models, on one NVIDIA GB200 NVL72 rack, and that it listens and talks at the same time and runs locally on a MacBook Pro M4 Max or better. For now it is a waitlist and a safety-research preview, with no price or product plan on the page.