Block Dessa Voice Model: The Speech-Only AI That Learned by Listening

Key Takeaways

• Block announced Dessa on 23 October 2026, a full-duplex speech-to-speech model that the company says was trained from scratch on recorded human conversation only, with no text and no pretrained models.
• Full-duplex means Dessa listens and talks at the same time, so it can interrupt, backchannel with short responses such as “yeah” or “got it”, and express emotion, laughter, and sarcasm, according to Block.
• Block says Dessa runs faster than real time and can run locally on a MacBook Pro with an M4 Max chip or better, and that training used a single Nvidia GB200 NVL72 system.
• Dessa is a research preview, not a public product. Early access is limited to a small group of AI safety researchers, with a public waitlist open.
• What Block has not published matters as much as what it has claimed. There is no parameter count, dataset size, API, pricing, weights, or formal benchmark results yet, so independent verification is still to come.
• Early coverage describes some demo conversations as natural and others as awkward, which is a useful reminder that polished demos are not the same as proven everyday performance.

Block, the payments company led by co-founder and chairman Jack Dorsey, has unveiled a voice model that takes a deliberately different path from most modern AI. Instead of starting from a large language model trained on text and then adding a voice layer, Block says its new model, Dessa, learned language by listening. The announcement, made on 23 October 2026, describes Dessa as a true speech-to-speech model: audio goes in, audio comes out, and no transcription step sits in between. That design choice is the whole story, because it changes what the model can hear, what it can express, and what it still needs to prove.

What Block actually announced

Dessa comes from a small AI research lab inside Block, led by Ragavan Thurairatnam and reporting to Dorsey. The lab says it was founded by former members of Dessa, a Toronto deep-learning startup, and that its team has been building generative speech models for more than eight years. Block calls Dessa the most authentic full-duplex speech model ever built, and Dorsey has described talking with it as feeling like talking with a person. Those are company claims, and they should be read as such. They describe an ambition and a first impression, not a measured result.

The concrete facts are narrower, and they are still interesting. Dessa is a research preview. Block has released demo recordings, including one conversation with OpenAI’s GPT-Live-1, and opened a public waitlist. Early access is limited to a small group of AI safety researchers. There is no public API, no published pricing, and no downloadable model yet. Anyone who wants to try Dessa today should expect to wait, and should treat third-party claims of access with caution until Block confirms them.

Why speech-only training is a big claim

Most voice assistants work in stages. They convert speech to text, send the text to a language model, generate a text reply, and then convert that reply back to speech. That pipeline works, but it throws away information at every step. Pauses, breathing, laughter, hesitation, tone, and timing do not survive transcription cleanly, and the result can sound correct but flat.

Dessa is designed to avoid that loss. Because it was trained, according to Block, only on recorded human conversation, it had to learn language from sound itself, including accents, pauses, breathing patterns, humour, sarcasm, inflection, and emotion. Block also says it learned when to stop talking, when to cut in, and when to backchannel, the short signals people use to show they are following along. In principle, that should produce conversation that feels more natural, because the model is modelling conversation itself rather than modelling text about conversation.

There is a second claim tucked inside the first. Block says Dessa learned language from hundreds of times fewer words than large language models see. That is a striking statement about data efficiency, but it needs context. A single spoken word consumes far more data than the same word in text, because audio carries pitch, timing, and voice quality as well as the word itself. Block says the team found a way to train Dessa on a single Nvidia GB200 NVL72 system, which is powerful hardware, but far smaller than the clusters used to train frontier text models. If independent testing supports that efficiency claim, it would matter for research teams that cannot afford frontier-scale compute. Until then, it remains a company claim without published training details to check against.

What full-duplex conversation feels like

Full-duplex is the technical term for listening and talking at the same time, and it is what separates Dessa from a system that waits politely for you to finish before replying. In a real conversation, people overlap, interrupt, laugh mid-sentence, and signal attention while someone else is speaking. Block says Dessa can do those things, and that its inference runs faster than real time, so replies arrive without an awkward pause.

The demos released so far give a mixed picture, and that mix is informative. Coverage of the announcement notes that the demos are mostly natural but occasionally awkward, with one example in which the model mistakes the speaker for someone named Ronaldo. The Verge described some sample conversations as a little awkward and uncanny. That is not a failure of the idea. It is a sign of how hard the problem is. Human conversation is full of ambiguity, repair, and social judgement, and a model trained only on speech still has to learn all of it from sound. The awkward moments also help explain why Block is starting with safety researchers rather than a general release. A model that can interrupt, imitate emotion, and sound convincingly human raises questions about impersonation, manipulation, and trust that are better examined before wide deployment.

For readers who are still building their understanding of how AI systems work, our beginner’s guide to learning AI in 2026 explains the basics of models, training data, and evaluation in plain language. Dessa is a good case study for that guide’s central lesson: a model’s training data shapes what it can do, and changing the data changes the system in ways that are not always obvious from a demo.

What Dessa is not, at least not yet

It is easy to overread an announcement like this, so it is worth stating plainly what Dessa is not. It is not a replacement for a voice assistant, a customer-service agent, or a phone product. Block has not announced a product built on Dessa, a release date, a price, or an integration with Square or Cash App. It is not an open model, because no weights have been published. It is not a benchmarked leader, because no formal benchmark results have been published. And it is not proof that speech-only training beats text-based training in general. It is one research result, presented through company demos, that suggests speech-only training can produce strikingly natural conversation in at least some cases.

That distinction matters for businesses considering voice AI. A research preview is a signal about where the field may go, not a tool to build on today. Teams evaluating voice products should keep asking the practical questions: how does the system handle accents, background noise, and overlapping speech in real calls, how is it evaluated, what safeguards exist against misuse, and what happens when it misunderstands a speaker. Dessa’s announcement does not answer those questions yet, and honest reporting should not pretend otherwise.

Why Block is doing this

Block is best known for payments, through Square and Cash App, so a speech research lab may seem like an odd fit. There are at least two plausible readings, and both may be true. The first is that voice is becoming a general interface, and a company that serves sellers and consumers has a long-term interest in how people talk to machines. The second is that Dorsey has a long-standing interest in open, decentralised technology and in research that challenges the dominant approach, and a small lab reporting directly to him fits that pattern. Neither reading is confirmed by Block as a product strategy, and this article does not claim otherwise. What is confirmed is the lab’s existence, its leadership, and its research direction.

What to watch next

The next steps are clear, and readers can judge Dessa by whether they happen. First, will Block publish a technical report with dataset details, model size, latency figures, and evaluation methods. Second, will independent researchers who get early access publish their own assessments, including failures as well as successes. Third, will Block announce an API, a product, or a timeline, or will Dessa remain a research project. Fourth, will the lab address data sourcing and safety questions in detail, given that a highly natural voice model carries clear misuse risks.

For now, the fair summary is this: Block has announced a genuinely interesting research model with a distinctive training approach, shown through demos that are often natural and sometimes awkward, and released it only to a small group of safety researchers behind a waitlist. That is enough to be excited about, and enough to stay sceptical about, at the same time. The rest of the story depends on evidence Block has not yet published.

If you want to think through how voice AI could fit a real business, rather than just a demo, our guide to how AI can help your business walks through seven workflows that work in practice, and where to be cautious before adopting a new tool. Dessa is not on that list yet, and that is the point. Promising research becomes a useful tool only when it survives contact with real users, real noise, and real evaluation.

Sources: Block unveils Dessa, a full-duplex voice model trained only on human speech and Jack Dorsey’s Block unveils Dessa, a full-duplex voice AI trained only on human speech and Jack Dorsey’s Block unveils new voice AI model that talks like a real human