Block unveils Dessa, a full-duplex speech model it says was trained only on human speech

Block, the company behind Square and Cash App, has unveiled Dessa, a research-preview voice model that it says learned to hold a live conversation from human speech alone, with no text and no pretrained models. Block announced it on 23 October 2026. There are no weights, no benchmarks and no public demo to try yet, so for now everything about it is the company's own account, backed by recordings it chose to publish.

Key points

• Block announced Dessa on 23 October 2026 as a full-duplex speech-to-speech model: it listens and talks at the same time.
• Block says it was trained from scratch on "nothing but human conversation": no text and no pretrained models.
• The company says it trained Dessa on a single NVIDIA GB200 NVL72 and that it runs faster than real time, locally, on a MacBook Pro M4 Max or better.
• Model size, training-data volume, licence and benchmark results are not stated; access is a waitlist, with early access for a small group of AI safety researchers.
• Block itself qualifies its "first" claim with "as far as we know", and the demo recordings carry subtitles added afterwards.

What happened

Block published a page for Dessa and a set of audio recordings of people talking to it. The Verge's Jay Peters, who covered the announcement the same day, described the chats as mostly natural back-and-forth that can also be a little awkward and uncanny. In one clip, he notes, the model at first mistakes the speaker for someone called "Ronaldo". Block's page lists others: a conversation about Christopher Nolan's films, a person telling Dessa about a surprise trip, a conversation about remembering a movie title, and one between Dessa and OpenAI's GPT-Live-1, a third-party voice model.

Block's page makes several specific claims about what the model does:

• It is "a true speech-to-speech model": audio in, audio out, with no step that turns speech into text.
• It is full-duplex. It listens and talks at the same time, and Block says it learned from real conversation when to cut in and when to give short acknowledgements such as "yeah" or "got it".
• It hears and produces emotion, sarcasm, laughter and sighs natively, which Block argues text-based pipelines throw away.
• Inference runs faster than real time and the model can run locally on a MacBook Pro M4 Max or better.
• Training used one NVIDIA GB200 NVL72, and Block says Dessa learned language from "hundreds of times fewer words" than large language models.

The last claim is a comparison, not a number: Block does not say how many hours or words of speech were used.

On access, Block runs a public waitlist and an application form for researchers (it asks for a university, company or lab email). The team says it will give "a small group of AI safety researchers" early access to look for risks it has not anticipated. Block's page says nothing about releasing the weights, and gives no licence.

Block says the demo recordings are live recordings of a person talking to Dessa, that nobody steers the model, and that subtitles were added afterwards. In the GPT-Live-1 demo, both voices are AI-generated, and Block states that GPT-Live-1's developer is not affiliated with Block and has not reviewed or endorsed Dessa. The Verge notes that in that clip Dessa comes across as the more natural of the two; that is a reporter's impression of a company-selected recording, not a measurement.

Background

Most voice assistants are pipelines. A speech recogniser turns your voice into text, a language model writes a reply, and a speech synthesiser reads it out. That design reuses powerful text models but, as Block points out, discards everything in a voice that text cannot carry: tone, hesitation, laughter, how loud or fast someone speaks. It also tends to wait for you to finish before answering.

Research on full-duplex spoken dialogue, where the system can listen and speak simultaneously, has been growing. A 2025 survey on arXiv ("From Turn-Taking to Synchronous Dialogue: A Survey of Full-Duplex Spoken Language Models") separates systems that coordinate listening and speaking with engineered modules from those that learn the synchronisation end to end. Block says Dessa learned to cut in and give acknowledgements from real conversation, which is closer to the second approach, though Block has not published the architecture and that placement is our reading.

Block's team is also not new to this. Its page says the lab is a small group at Block founded by former members of Dessa, the Toronto deep-learning startup, that it has built deep-learning models since 2012 and generative speech models for over eight years, and that it reports directly to Jack Dorsey, Block's co-founder and chairman. Ragavan Thurairatnam is the lead, with seven named researchers, three interns and two former interns.

How it works

Block does not publish an architecture or a paper with this announcement, so we cannot describe how Dessa works internally. What the page offers is a design philosophy and a few numbers that frame the difficulty:

• Text is the shortcut Dessa skips. Block says large language models need an "insane amount of text" and that Dessa had to learn accents, pauses, breathing, sarcasm and when to interject on top of language.
• Speech data is heavier. Block says a single spoken word takes thousands of times more bytes than the same word in text, and that speech never repeats exactly: saying the same sentence twice yields different data.
• Same words, different meaning. The page's illustration is stressing different words in "I didn't say she stole my CUDA code" to change what is meant.
• Training budget. One GB200 NVL72 is a single rack-scale NVIDIA system. Block presents the modest hardware as part of the achievement, but we cannot compare it with other labs' training runs because no run sizes are given.

What it means

If the claims hold up, Dessa is interesting less as a product than as a result: a conversational model that never saw text could be a new kind of research object. Block itself asks questions such as how efficiently such models learn language and facts compared with text, how efficiently they learn new languages, and what in-context learning they can do. Those are open questions that only outside access and a technical report can answer.

For developers, the practical news is limited today. There is no API, no download and no licence, so nothing can be built on Dessa yet. The local-inference claim, on a high-end laptop, hints at a direction in which voice agents run on-device, with lower latency and no audio sent to a server, but that is Block's claim about its own system.

For the safety argument, Block makes its case plainly: a voice AI that cannot hear distress "just keeps going", and a model that hears emotion could respond better. It also admits that a model that learned to talk from sound alone is "new territory", which is why it wants outside researchers to probe it. The same capabilities, convincing emotion and timing, are the ones that make synthetic voices easier to misuse, and the announcement does not describe safeguards.

For the competitive picture, the demo against GPT-Live-1 is a marketing choice as much as evidence. A head-to-head between two AI voices, picked and subtitled by one side, says little about how the models compare for ordinary users.

What we don't know yet

• Model size and architecture. Neither is stated.
• How much speech it was trained on, and where it came from. Block gives a comparison to LLM word counts but no figure, and does not say where the conversations came from.
• Quality beyond the demos. No benchmarks, latency figures or independent evaluations have been published, and subtitles in the demos were added afterwards.
• Whether and when it will be released. There is a waitlist, a researcher programme and a note that launch links are coming, but no release date, weights or licence.
• Safeguards. The announcement talks about safety research but does not describe protections against voice impersonation or misuse.
• Whether it is really the first. Block says "as far as we know" Dessa is the first realistic live-conversation AI trained only on human speech; we have not verified that.

What to watch next

Watch for a technical report or paper from Block's lab, any release of weights or an API, independent tests by the researchers who get early access, and responses from the large voice-AI vendors. Block says more recordings will be shared on X. The hardware claim, a laptop-class machine running the model faster than real time, is the easiest to check once anyone outside Block can run it.

Sources

• Dessa, the conversation model trained only on speech (Block, primary source)
• You have got to listen to these conversations with Block's new Dessa AI model (The Verge)
• Block Unveils Dessa, a Full-Duplex Speech Model Trained on Conversation (Unite.AI)
• From Turn-Taking to Synchronous Dialogue: A Survey of Full-Duplex Spoken Language Models (arXiv)