Block Unveils Dessa, a Full-Duplex Speech Model Trained Only on Human Conversation

Block on October 23, 2026 unveiled Dessa, a speech model that its research lab says was trained from scratch on nothing but recorded human conversation. Block says Dessa can hold a live, real time conversation and runs locally on a MacBook Pro M4 Max or better, but there is no public access yet.

The announcement puts Jack Dorsey's payments company, the owner of Square and Cash App, among the builders of new AI models for voice, a field where OpenAI, Kyutai and Hume AI have shipped conversational systems since 2024 (The Decoder). Block published more than 20 recorded conversations with the model on its Dessa page instead of benchmark scores.

What is Dessa?

Dessa is a full-duplex speech-to-speech model: audio goes in, audio comes out, and it listens and talks at the same time, Block said on October 23, 2026. Block says everything the model knows, from words to facts, came from human speech, with no text and no pretrained models in its training.

Most voice assistants turn speech into text first, which Block says throws away tone, sarcasm, laughter and sighs; Dessa hears and produces those cues natively. Because it learned from real conversation, Block says, Dessa knows when to cut in and when to backchannel, the short "yeah" or "got it" sounds listeners make to show they are following. Block also says inference runs faster than real time.

Block claims Dessa learned language from "hundreds of times fewer words" than large language models need, and that the team trained it on a single NVIDIA GB200 NVL72, one rack of NVIDIA's data center GPUs. The company calls Dessa "the most authentic full-duplex speech model ever built" and, as far as it knows, the first AI trained only on human speech that can hold a realistic live conversation. Block cites no independent test for either claim.

What do the demos show?

Block's demos are live recordings of people talking with Dessa, with subtitles added afterward, and the company says nobody steers what the model says or how it sounds. One clip pairs Dessa with OpenAI's GPT-Live-1, whose developer, Block notes, has not reviewed or endorsed Dessa.

The Verge described some exchanges as awkward and uncanny: in one, the model mistakes the speaker for someone named "Ronaldo". The Verge also wrote that Dessa comes across as more natural than GPT-Live-1 in their exchange. TechCrunch noted it is not clear whether Dessa's emotion is steered by tags, the markup that ElevenLabs v4 and Fish Audio Drama 3 use to direct emotional delivery.

Can anyone use Dessa yet?

No: as of October 24, 2026, Dessa is not publicly available, and the download and waitlist buttons on Block's page only show messages that they are coming soon. TechCrunch's report says Block "released" the model, but the same article states there is no public access at the moment, which matches The Decoder's report.

The only access on offer is for AI safety researchers. Block says it is giving a small group of them early access to find risks it has not anticipated; applicants sign up with an institutional email address. Block's page says nothing about model weights, a license, pricing or an API.

How does Dessa compare with GPT-Live and Moshi?

Dessa's main difference from the best known full-duplex systems is the absence of text: Kyutai's Moshi and OpenAI's GPT-Live both lean on a text trained language model, while Block says Dessa has none. The details below come from Kyutai's Moshi paper, which calls its text step Inner Monologue, and from SiliconANGLE's report on the GPT-Live launch.

The trade off Block is betting on is that a speech only model keeps information text loses. The Decoder argues the other side: sounding emotional is not the same as being intelligent, and for advanced voice AI the ability to reason while responding matters more than a natural voice.

Who built Dessa, and why does Block want it?

Dessa comes from a very small AI research lab inside Block that was founded by former members of Dessa, the Toronto deep learning startup, according to the announcement. The lab reports directly to Jack Dorsey, Block's cofounder and chairman. Square, the Block business, said in February 2020 that it had acquired Dessa, a company then known for its work on deepfake detection (Square); BetaKit reported at the time that Dessa would keep its headquarters in Toronto (BetaKit).

The project is led by Ragavan Thurairatnam. Block gives three reasons for the work: curiosity, better understanding between humans and AI, and safety, arguing that a voice AI that cannot hear distress in a user's voice is a risk. TechCrunch points to a commercial angle too: Square already uses voice AI for ordering, and a voice model that runs on device could power more business and consumer features at lower cost.

What we don't know yet

• When the waitlist and the download will open, and whether the download includes the model weights and under what license.
• What Dessa will cost, and whether it will be offered through an API or inside Square and Cash App products.
• How Dessa scores on independent tests of accuracy, latency and conversation quality; Block has published recordings, not benchmarks.
• Which languages Dessa supports and how much speech data it was trained on.
• What the safety researchers with early access find.