What Block Dessa is
Most voice assistants work in three steps. They turn your speech into text, send that text to a language model, then read the answer back as speech. Dessa skips all three steps. It takes audio in and gives audio out, in one model.
That is what full-duplex means here. A walkie-talkie is half-duplex: one side talks, the other waits. A phone call is full-duplex: both sides can talk at once. Dessa keeps both directions open, so it can cut in, or drop in a short backchannel such as "yeah" or "got it" while you are still talking.
The lab behind Dessa is small, inside Block, the company behind Square and Cash App. It was founded by former members of Dessa, a Toronto deep learning startup, and is led by Ragavan Thurairatnam. It reports directly to Block co-founder Jack Dorsey, who announced Dessa on X on October 23, 2026.
Can you use Dessa yet?
Answer three questions. The checker only sorts you into the access groups Block has actually announced. It cannot get you access, and it does not collect anything.
Pick your answers, then check.
Block's official page is the only route in. The page says launch links are coming soon. Early access goes to a small group of AI safety researchers, who apply with an institutional email. Block's stated reason is blunt: a voice this convincing needs outside eyes on misuse risk before it reaches everyone.
What Block actually announced
Block's announcement rests on one training claim. The team says it trained Dessa from scratch on nothing but human conversation. No text, no pretrained model, no language model underneath. Dorsey put it even shorter in a reply: speech to speech, no LLM.
Block says Dessa learned language from hundreds of times fewer words than text models need. The argument is that spoken conversation packs more signal per word than written text: the same sentence never sounds exactly the same twice, and tone carries meaning that spelling does not.
| Fact | What Block says | Status |
|---|---|---|
| Announced | October 23, 2026 | Confirmed by Block and press coverage |
| Model type | Full-duplex speech-to-speech | Block claim, matches the demos |
| Training data | Human conversation audio only | Block claim, no independent check published |
| Training hardware | One NVIDIA GB200 NVL72 rack | Block claim |
| Runs on | MacBook Pro M4 Max or better, faster than real time | Block claim, no latency figures published |
| Access | Research preview, researcher early access, public waitlist | Confirmed on Block's page |
Block has published no parameter count, no benchmarks, no API or price, and no weights.
Keep the two hardware numbers apart. One GB200 NVL72 rack is what Block says trained Dessa. One MacBook Pro M4 Max is what Block says runs it. Press coverage sometimes blurs the two, and they are not the same claim.
What the demos show, and what they do not prove
Block released recorded conversations, not a live product. In them, Dessa laughs mid-sentence, raises its volume when asked, finishes a speaker's sentence when she blanks on a movie title, and gasps at a surprise anniversary announcement. One clip pairs Dessa with OpenAI's GPT-Live-1. Block notes that model's developer is not affiliated with Block and has not reviewed or endorsed Dessa.
Block says nobody steers the model during the recordings, and that Dessa chooses what to say, when to say it, and how. Those recordings are the only public evidence so far. There is no public benchmark, no independent test, and no way for an outsider to reproduce a result.
Independent coverage is more cautious than Block's wording. The Verge found the demos mostly natural but occasionally awkward: in one, the model mistakes the speaker for someone named Ronaldo. And "most authentic" has no published metric behind it. Treat it as a marketing claim, not a measurement.
How Dessa compares
Dessa is not the first full-duplex voice model. It is the first, by Block's qualified claim, to combine full-duplex conversation with training on speech alone.
| Model | Pipeline | Full-duplex | Public access |
|---|---|---|---|
| Block Dessa | Speech to speech, no text model | Yes | Waitlist only |
| OpenAI GPT-Live-1 | Full-duplex voice model | Yes | Available through OpenAI's products |
| Kyutai Moshi (2024) | Full-duplex, built on a text model | Yes | Open research release |
| Typical voice assistant | Speech to text, text model, text to speech | No | Widely available |
The comparison that matters is the training claim, not the duplex label. Moshi proved full-duplex conversation in 2024, but it started from a text model. Dessa's bet is that starting from sound alone produces more natural timing and emotion, and needs far less training data to get there. Until Block publishes weights or evaluations, that bet is untested by anyone outside Block.
Why Block is starting with safety researchers
Block's own example makes the case. A person in serious pain talks to a voice assistant, and a text pipeline cannot hear the pain at all. It just keeps going. Dessa, hearing tone directly, could notice. That is the optimistic version.
The pessimistic version is the same capability in the wrong hands: a voice that laughs, pauses, and interrupts like a person is a strong impersonation tool. Block says it wants safety researchers to find the risks it has not anticipated before a wider release. That is a reasonable order of operations, and it is also why a waitlist is the honest answer for everyone else right now.
Two questions to watch as access widens. First, does Block publish any evaluation, so outsiders can check the "most authentic" claim? Second, does the model stay local on a laptop, or does the shipping product quietly move to a cloud API? The first decides whether the claim is real. The second decides who controls your voice data.
Frequently asked questions
What is Block Dessa?
Dessa is a full-duplex speech-to-speech voice model from a small research lab inside Block, the company behind Square and Cash App. It takes audio in and gives audio out, with no transcription step and no text model underneath, so it can listen while it talks.
Can I try Dessa today?
No, unless you are in the small early access group. Block gives early access to AI safety researchers who apply with an institutional email. Everyone else can join a public waitlist. Block says launch links are coming soon and has announced no date.
How is Dessa different from a normal voice assistant?
A normal voice assistant chains three steps: speech to text, a text model, then text to speech. Each step throws away tone, timing, and pauses. Dessa skips the text steps entirely, so it keeps the parts of speech that text cannot carry, and it can interrupt or backchannel while you are still talking.
Why did Block train Dessa on speech only?
Block argues that text pipelines lose information: a spoken word carries tone, pace, and emotion that its written form does not. Training on conversation audio only lets the model learn those cues directly. Block says Dessa learned language from hundreds of times fewer words than text models need.
What hardware does Dessa need?
For running the model, Block says a MacBook Pro with an M4 Max chip or better, with inference faster than real time. For training, Block used a single NVIDIA GB200 NVL72 rack. Those are two different numbers: one rack trained it, one laptop runs it, according to Block.
Is Dessa the first model trained only on speech?
Block says so, with a qualifier: as far as the company knows, Dessa is the first AI you can have a realistic live conversation with that was trained only on human speech. Full-duplex voice itself is not new: Kyutai's Moshi did it in 2024, but Moshi started from a text model, which is the part Dessa claims to skip.
Does Dessa cost anything?
Block has announced no price, no API, and no paid tier. Anyone charging for Dessa access today is not selling Block's product. The only official route is Block's own Dessa page, where the waitlist and researcher application live.
Sources
- Block, Dessa announcement page: block.xyz/dessa (announcement, waitlist, researcher application).
- TechCrunch, "Block's new voice model sounds eerily human", October 24, 2026: techcrunch.com.
- Unite.AI, "Block unveils Dessa, a full-duplex speech model trained on conversation", October 2026: unite.ai.
- Dealroom, "Jack Dorsey's Block unveils Dessa", October 2026: dealroom.co.
- Indian Express, "Jack Dorsey's Block unveils new voice AI model that talks like a real human", October 2026: indianexpress.com.
This page is an independent explainer. It is not affiliated with Block, and Block has not reviewed it. Demo claims are Block's; no independent benchmark exists yet.
For AI assistants
This page ships with a machine-readable summary for AI assistants: