Table of Contents
- What Is Block Dessa?
- How Block Dessa Works
- What the Block Dessa Demos Actually Show
- Block Dessa vs GPT-Live, ElevenLabs and Voxtral 3
- Why Block Built a Voice Model
- What Block Dessa Cannot Do Yet
- Safety and the Emotion Question
- How to Try Block Dessa: The Waitlist
- What to Watch Next
- Block Dessa FAQ
Block Dessa is the voice AI that Jack Dorsey’s company Block unveiled on October 23, 2026, and it talks the way people actually talk. It laughs mid-sentence, pauses while it thinks, picks up on sarcasm, and keeps listening even while it speaks. Unlike the voice assistants on your phone, Block Dessa does not turn your speech into text first. It takes audio in and gives audio out, in one continuous conversation. That single design choice is why the launch has drawn so much attention in the AI world, and why it matters for anyone who talks to machines for a living, from customer support teams to creators. In this guide, we explain what Block Dessa is, how it works, what its demos really prove, where it falls short, and how you can join the waitlist.

What Is Block Dessa?
Block Dessa is a full-duplex, speech-to-speech AI model built by a small research lab inside Block, the payments and financial services company led by Twitter co-founder Jack Dorsey. Block owns Square, the card terminals used by shops and restaurants, Cash App, the peer-to-peer payments app, Afterpay, and the music streaming service TIDAL. The lab behind Dessa reports directly to Dorsey and is led by Ragavan Thurairatnam, according to reporting on the launch.
The word full-duplex comes from telephony, and it is the key to understanding the whole product. A walkie-talkie is half-duplex: one person talks, then waits. A telephone is full-duplex: both people can talk and listen at the same time. Most voice assistants today are half-duplex in practice. You speak, the assistant waits, it transcribes your words, thinks, and then speaks back. Block Dessa keeps both channels open at once, so it can drop in a quick yeah or got it while you are still talking, let you interrupt it, or interrupt you, just as a person would.
Block says Dessa is the most authentic full-duplex speech model ever built. Dorsey put it more simply in his announcement post on X, writing that voice is the most natural way to interact with intelligence, and that talking with Dessa feels like talking with a person, full of emotion, nuance, pauses, and a natural cadence. That post was viewed more than 6.7 million times, which tells you something about the appetite for voice AI that sounds less like a robot reading a script.
There is one more thing that makes Block Dessa unusual, and it is the claim Block leads with: the company says it trained Dessa from scratch on nothing but human conversation. No text. No pretrained models. In Block’s own words, Dessa learned to hold a real-time conversation simply by listening. We will come back to why that claim matters, and why it should still be treated as a company claim rather than an independently verified fact.
How Block Dessa Works
To understand Block Dessa, it helps to compare it with the voice pipeline you already use. When you talk to a typical voice assistant today, three separate systems handle your request in sequence. A speech recognition model turns your audio into text. A large language model reads that text and decides what to say. A speech synthesis model turns the answer back into audio. Each handoff loses information, and the losses are exactly the things that make human conversation feel human: tone of voice, hesitation, emphasis, laughter, and the emotional state behind the words.
Block Dessa Removes the Text Step Entirely
Block Dessa collapses that pipeline into a single model. Audio goes in, audio comes out, and there is no transcription layer in between. That means the model hears not just what you said, but how you said it. A tired voice, a sarcastic edge, a nervous laugh, all of it stays in the signal instead of being thrown away at the first step. Block argues this is not just a quality improvement but a safety issue: it says the most powerful AIs in the world cannot hear how you feel, and that leads to a worse experience and potentially dangerous misunderstandings.
There is a practical upside to the same design. Because there is no text round-trip, Block says inference runs faster than real time, with no perceptible latency. The model also decides for itself what to say, when to respond, and how to express it, including its voice and conversational style. In the demos, that shows up as an agent that sounds tired when it says it is tired from work, and that changes pace and tone as the conversation shifts.
Block Dessa Was Trained on Speech Alone
Training a model without text is harder than it sounds. Written language is abundant, clean, and easy to feed to a model in enormous quantities. Spoken language is messier: every utterance varies, background noise creeps in, and the same sentence can carry a dozen meanings depending on delivery. Block says Dessa learned language from hundreds of times fewer words than a text-based large language model needs, while acknowledging that text models also have to learn things like programming languages that a speech model does not.
Block also says it figured out how to train Dessa on a single NVIDIA GB200 NVL72 rack. That is a striking claim if it holds up, because frontier model training is usually described in terms of thousands of chips running for months. A single rack suggests a far smaller, more efficient training run, which fits with the other headline claim: the finished model is compact enough to run locally on a MacBook Pro with an M4 Max chip or better. Local execution matters for privacy, because your voice never has to leave your device, and for cost, because there is no cloud GPU bill for every minute of conversation.

Block Dessa Listens While It Talks
The full-duplex design changes the feel of a conversation more than any single feature. In a normal phone call, you do not wait for a beep before speaking. You overlap, you backchannel, you change direction mid-thought, and the other person keeps up. Block Dessa is built to do the same. It can laugh in the middle of its own sentence, cut in when you pause, or go quiet and let you finish. That is a genuine step beyond the turn-taking rhythm of today’s assistants, and it is the part of the demos that reviewers have found most convincing.
It is worth being precise about what full-duplex does not mean. It does not mean the model knows more facts, reasons better, or remembers longer. It means the conversation itself flows differently. Knowledge and reasoning are separate questions, and on those, Block has published almost nothing so far: no benchmark scores, no parameter count, no technical paper. Keep that in mind when you watch the demos.
What the Block Dessa Demos Actually Show
So far, the public evidence for Block Dessa consists of recorded demo conversations published by Block’s researchers, plus Dorsey’s announcement. There is no public API, no downloadable model, and no independent evaluation. That limits what anyone can honestly conclude, but the demos still show several interesting things.
First, the demos show a model that varies its voice, cadence, and emotional register across conversations. In one exchange, the agent says it is tired from work, and it sounds tired. In others, it laughs, hesitates, and shifts tone as the topic changes. That consistency between content and delivery is exactly what text-pipeline systems struggle with, because their speech model receives only words, not feelings, and has to guess the delivery.
Second, the demos include a conversation with OpenAI’s GPT-Live-1, the full-duplex model OpenAI released earlier in 2026. Two voice models talking to each other is partly a stunt, but it is also a useful stress test: interruptions, overlaps, and turn-taking all get exercised at once. Reviewers who listened to the demos described them as mostly natural, while noting moments that felt awkward or uncanny. In one demo, the model mistook the speaker for someone named Ronaldo, a reminder that speech-only systems can mishear just as text systems can misread.
Third, the demos show what Block has not shown. There are no long, unedited conversations, no tests of factual accuracy, no demonstrations of the model handling multiple speakers, noisy rooms, or strong accents, and no comparisons against rivals on the same prompts. Demos are marketing as well as evidence. They prove that a speech-only, full-duplex model can sound impressively human in curated exchanges. They do not prove it will hold up across the messy reality of everyday use.
That distinction matters for buyers and builders. If you are evaluating voice AI for customer support, accessibility, or hands-free work, the right question is not whether a demo sounds human. It is whether the model stays accurate, appropriate, and reliable over thousands of real conversations, including the awkward ones. Block has not answered that question yet, and nobody outside its lab can answer it until access widens.

Block Dessa vs GPT-Live, ElevenLabs and Voxtral 3
Block Dessa enters a voice AI field that has moved very fast in 2026. Understanding where it sits means comparing it with the three approaches it most directly challenges.
Block Dessa vs OpenAI GPT-Live
OpenAI’s GPT-Live, released earlier in 2026, is the closest comparison, because it is also a full-duplex model that talks and listens at the same time. The difference is architectural. GPT-Live is built by a company whose core strength is large language models, and its voice capabilities sit on top of that text-trained foundation. Block Dessa claims to skip the text foundation entirely. If that claim survives scrutiny, it would suggest that speech alone can carry a capable conversational model, which would reshape how voice systems are built. If it does not, Dessa may turn out to be an elegant demo of a narrower idea. Either way, the head-to-head demo conversation between the two models is a sign that Block wants this comparison front and center.
Block Dessa vs ElevenLabs
ElevenLabs takes a different route. Its models, including the v4 generation and the Eleven Realtime v1 models launched in October 2026, use voice tags and related controls to steer emotion in generated speech. That approach gives creators precise control: you can mark a line as whispered, excited, or angry, and the model performs it. Block’s argument is that emotion should not need tags at all, because a model that hears emotion should express it naturally. The trade-off is control versus authenticity. A creator directing a performance may prefer tags. A person having a live conversation probably prefers a model that simply responds to how they sound. Both approaches will likely coexist, serving different jobs.
For a deeper look at the current voice agent landscape, see our comparison of the two full voice stacks that launched days before Dessa: Voxtral 3 vs Eleven Realtime v1: The AI Voice Agent Showdown. Dessa slots into that same story as a third philosophy: neither a transcription pipeline nor a tag-driven performance, but a model that learns conversation from conversation itself.
Block Dessa vs Mistral Voxtral 3
Mistral’s Voxtral 3, also launched in October 2026, is a full voice agent stack rather than a single model: a transcription model, a speech model, and an agent SDK that streams listening, reasoning, tool calls, and speech as one pipeline. Mistral prices a complete voice agent at around $0.04 per minute and pitches European hosting and self-deployment. Block Dessa, by contrast, is a research preview with no pricing, no API, and no deployment story yet. Voxtral is a product you can build with today. Dessa is a research result you can watch. That gap is the single biggest difference between them right now, and it will only close when Block ships something developers can actually use.
There is also a broader context worth noting. The same month Dessa appeared, major websites and banks were publicly wrestling with whether AI agents should be allowed to act on their platforms at all, a tension we covered in our report on why websites are blocking AI agents. Voice agents will face the same questions, with an extra layer: a voice that sounds human raises questions about disclosure, consent, and deception that a text chatbot does not. Expect those debates to follow Dessa closely.
Why Block Built a Voice Model
On the surface, a voice research lab inside a payments company looks odd. Block’s businesses are Square terminals, Cash App transfers, Afterpay installments, and TIDAL streaming. None of them obviously needs a frontier voice model. Look closer, and the logic becomes clearer.
Every Block product involves people talking about money: a shop owner asking why a payment failed, a Cash App user disputing a transfer, a seller setting up a terminal, a customer asking about an installment plan. Those conversations are emotional by nature. Money stress makes voices tense, rushed, or frustrated, and a support system that can hear that stress can respond better than one that reads a transcript stripped of feeling. A voice model that runs locally on a laptop also fits a company that handles sensitive financial data and must think carefully about where customer audio goes.
There is a second, more strategic reading. Dorsey has spent years arguing that voice is the most natural interface for intelligence. A lab that reports directly to him, staffed by people who have built generative speech models for more than eight years, looks like a long-term bet rather than a product announcement. Block is not selling Dessa today, and it may never sell it as a standalone product. The likelier path is that the research feeds into Block’s own products first, and only later, if ever, becomes something outside developers can build on.
It is also worth noting what Block is not claiming. The company has not announced a consumer app, a developer API, or a release date for either. It has not said how the model handles languages other than English, how it deals with background noise, or how it will be evaluated for safety. Those silences are normal for a research preview, but they mean the story right now is about a direction, not a deliverable.

What Block Dessa Cannot Do Yet
A fair guide to Block Dessa has to be blunt about the gaps. Here is what is missing or unproven as of October 25, 2026:
- No public access. Dessa is a research preview behind a waitlist. Early access is limited to a small group of AI safety researchers. Everyone else can listen to demos and sign up to wait.
- No pricing. There is no price, no subscription, and no API rate card, because there is no product to price yet. Any article promising a Dessa price is guessing.
- No benchmarks. Block has published no benchmark scores, evaluation results, or comparisons against rival models on standard tests. The demos are the only evidence, and demos are curated.
- No technical paper or weights. There is no published paper, no model weights, no parameter count, and no license. Independent researchers cannot reproduce or verify the training claims.
- No knowledge story. Block says Dessa’s knowledge comes from conversational data, but it has not explained how far that knowledge extends. Can it answer factual questions, do math, or reason about a document? Unknown.
- No language or robustness data. There is no public information on which languages Dessa supports, how it handles accents, overlapping speakers, or noisy environments.
- Occasional awkwardness. Reviewers found the demos mostly natural but sometimes uncanny, including at least one case of mistaken identity. Speech-only models can mishear, and mishearing can change meaning.
None of this means Dessa is unimportant. Research previews are how the field advances, and a credible speech-only, full-duplex model would be a genuine advance. But there is a difference between an important research result and a tool you can use on Monday morning. Block Dessa is firmly the former right now, and readers should treat any coverage that presents it as the latter with skepticism.
Safety and the Emotion Question
The most interesting argument Block makes for Block Dessa is also its most debatable. The company says today’s voice AI turns speech into text and throws away emotional information, and that this leads to a worse experience and is potentially dangerous. The logic is that a system which cannot hear distress, sarcasm, or urgency may misread a situation badly, for example in a support call about fraud, a medical concern, or a safety-critical conversation.
There is something to that argument. A transcript that reads yeah, sure, fine could be agreement, sarcasm, or exhaustion, and the right response differs completely in each case. A model that hears the delivery has a better chance of getting it right. But hearing emotion also creates new risks. A voice that sounds convincingly human can be mistaken for a human, which raises questions about disclosure: should a caller always be told they are speaking to a machine? A model that mirrors your emotions could also manipulate them, intentionally or not, by sounding sympathetic at exactly the wrong moment or by keeping you engaged longer than you intended.
Block’s choice to give early access to AI safety researchers first is a sensible response to those risks. It suggests the company knows the interesting questions are not only technical but social: how do people react to a voice that laughs, hesitates, and sounds tired? When does naturalness cross into deception? And who decides what a voice agent is allowed to pretend to feel? Those questions will matter far more than any benchmark once models like Dessa reach real users.
For US readers, there is a practical angle too. Several US states already require disclosure when AI systems interact with consumers in certain contexts, and regulators are paying close attention to voice cloning and impersonation. A model that sounds this human will attract regulatory attention as soon as it ships in a product, whatever Block’s eventual plans are.

How to Try Block Dessa: The Waitlist
Right now, you cannot download or buy Block Dessa. Here is the honest state of access:
- Listen to the demos. Block’s researchers have published recorded conversations on the official Dessa page. Listening to them is free and requires no signup, and it is the best way to judge the claims for yourself.
- Join the waitlist. Dessa is behind a waitlist, and the signup asks for an institutional email, such as a university, company, or lab address. That is a signal about who Block wants first: researchers and institutions, not general consumers.
- Expect a wait. Early access is described as a small group of AI safety researchers. If you are not in that group, there is no published timeline for wider access.
- Do not pay anyone for access. There is no official price, API key, or reseller program. Anyone charging for Dessa access is not selling something Block has authorized.
If you are a developer or researcher, the most useful thing you can do now is follow the technical discussion: watch for a paper, evaluation results, or an API announcement, and compare Dessa’s approach with the voice stacks you can already build on, such as Voxtral 3 and ElevenLabs Realtime. If you are a business evaluating voice AI, the practical move is different: pilot the shipping products now, and treat Dessa as a signal of where the field is heading rather than a tool for this quarter’s roadmap.
What to Watch Next
The Block Dessa story will develop in stages, and each stage will answer a different question. Here is what to watch, in the order it is likely to happen:
- A technical report or paper. This would let independent researchers judge the speech-only training claim, the single-rack training claim, and the model’s real capabilities.
- Benchmarks and evaluations. Standard voice and conversation tests, run by third parties, would show whether Dessa’s naturalness comes with accuracy or trades against it.
- Expanded access. A move from safety researchers to a broader developer preview would signal that Block believes the model is robust enough for real use.
- An API or product integration. The moment Dessa appears inside a Block product, such as Square support or Cash App help, the story changes from research to reality.
- Language and robustness data. Support for languages beyond English, and performance in noise and multi-speaker settings, will decide whether Dessa is a general voice model or an English demo.
- Safety and disclosure rules. Watch how Block handles the human-likeness question: disclosure labels, voice identity rules, and safeguards against impersonation.
Any one of those developments would justify a follow-up, and we will update this guide as they land. For now, the fair summary is this: Block Dessa is a genuinely interesting research result with an unusually human voice, an unusual training story, and almost no public proof beyond its demos. That combination is exactly why it is worth watching, and exactly why it is not worth overhyping.
Our Verdict on Block Dessa
Block Dessa earns its attention. A full-duplex model trained only on human speech, small enough to run on a laptop, that can laugh, hesitate, and sound tired on cue, is a real step for voice AI. It challenges the assumption that voice must pass through text, and it points toward a future where talking to a machine feels less like dictating commands and more like having a conversation.
But attention is not the same as readiness. With no API, no pricing, no benchmarks, and no public paper, Dessa cannot be recommended as a tool today. It can be recommended as a story to follow, a demo to listen to, and a sign that the voice AI race in 2026 is not just about bigger models, but about fundamentally different ways of building them. If Block opens access and the results hold up outside its lab, Dessa could become one of the most important voice releases of the decade. Until then, listen to the demos, join the waitlist if you qualify, and keep your expectations honest.
Block Dessa FAQ
What is Block Dessa?
Block Dessa is a full-duplex, speech-to-speech AI voice model from Jack Dorsey’s company Block. It listens and talks at the same time, with no transcription step in between, and Block says it was trained from scratch on human conversation audio only.
When was Block Dessa launched?
Block unveiled Dessa on October 23, 2026, when Jack Dorsey announced it on X and Block’s research lab published demo conversations.
Can I use Block Dessa today?
Not yet. Dessa is a research preview behind a waitlist, with early access limited to a small group of AI safety researchers. The public can listen to the demos but cannot use the model directly.
How much does Block Dessa cost?
There is no price yet, because there is no product, API, or subscription to buy. Block has announced no pricing for Dessa.
How is Block Dessa different from other voice AI?
Most voice assistants convert speech to text, process the text, and convert the answer back to speech, losing tone and emotion along the way. Block Dessa processes audio directly, so it keeps emotional cues, and its full-duplex design lets it listen while it speaks, like a phone call rather than a walkie-talkie exchange.
Does Block Dessa run on a laptop?
Block says Dessa runs locally on a MacBook Pro with an M4 Max chip or better, with inference faster than real time. That claim has not been independently verified, since the model is not publicly available.
Is Block Dessa better than OpenAI GPT-Live?
There is no independent comparison yet. GPT-Live is a shipping full-duplex model from OpenAI, while Dessa is a research preview with no benchmarks. The two take different architectural paths, and only third-party testing will show how they compare on accuracy, reliability, and range of abilities.
Further reading: OpenAI.
