Block unveils Dessa, a voice model trained only on human speech

What Block unveiled

Block, the company behind Square and Cash App, has shown off Dessa, a speech model it says was trained from scratch on nothing but human conversation. There was no text in the training data and no pretrained model underneath it. Block says it learned to hold a live conversation simply by listening, and calls it, as far as it knows, the first AI trained only on human speech that people can have a realistic real-time conversation with.

The model was publicised on 23 October 2026, when The Verge reported on the launch and the clips Block shared. The project page carries a privacy notice dated 21 October. Dessa is not publicly available: the page offers a waitlist, says launch links are coming soon, and invites AI safety researchers to apply for early access.

Block describes Dessa as the most authentic full-duplex speech model ever built. That is a company claim, not a benchmark result, and Block has not published a technical paper or evaluation numbers alongside it.

Audio in, audio out

Most voice assistants today work as a pipeline or are built on a text model. They turn speech into text, reason over the text, then turn a written answer back into speech. Even newer speech-to-speech systems usually start from a large language model that learned from written words.

Dessa is pitched as the opposite. Block calls it a true speech-to-speech model in which everything it knows came from human speech. Its argument is that converting voice to text throws away information that text cannot carry: emotion, sarcasm, laughter, a frustrated sigh or a voice cracking with sadness. Block says Dessa hears and produces those natively.

The company illustrates the point with a sentence whose meaning shifts depending on which word you stress. Spoken language carries that difference; a transcript does not. Researchers have been building datasets to teach models these cues, such as the Global Emotions collection of steered speech, but Block's claim is that Dessa picked them up from raw conversation alone.

Listening and talking at once

Full-duplex means the model listens and speaks at the same time, as people do on a phone call. Block says Dessa learned from real conversations when to cut in and when to backchannel with small signals such as yeah, got it or ooh to show it is following. It says inference runs faster than real time, so there is no awkward pause before it replies.

One example Block highlights is laughter. It says Dessa laughs the moment something is funny, even mid-sentence, while other models wait for the speaker to finish and then laugh on cue. Asked to speak up, it gets louder, which Block says nobody taught it.

Block also says the model runs locally on a MacBook Pro with an M4 Max chip or better. That would let it work without sending a conversation to a server, though Block has not said what download size or memory the model needs.

What the demos show

Block says every clip on its page is a live recording of a person talking to Dessa, with nobody steering it and subtitles added afterwards. Dessa picks what to say, when to say it and how, down to its own voice and personality.

The examples include a person blanking on a movie title while Dessa finishes the sentence, a conversation about Christopher Nolan films in which, as The Verge noted, it wanders off into Bryan Singer's Superman films, and a surprise anniversary trip to Paris that Dessa answers with a gasp and what Block calls teary gratitude. In one clip shared by the project lead, the model starts by mistaking the speaker for someone named Ronaldo.

Block also posted a conversation between Dessa and OpenAI's GPT-Live-1 voice model, with a note that OpenAI is not affiliated with Block and has not reviewed or endorsed Dessa. The Verge's reporter said Dessa came across as the more natural of the two. AI Lookout has covered OpenAI's GPT-Live updates and other recent voice launches, which all build on text-trained models.

The Verge described the conversations as occasionally a little awkward and uncanny, which matches what a listener might expect from a model that learned language without ever reading.

Learning language without text

Block's headline research claim is about efficiency. It says Dessa learned language from hundreds of times fewer words than large language models need, adding the joke that LLMs also have to learn C++ and the Angular web framework. The company did not give a precise word count or hours of audio.

It argues the task is harder than it looks. A sentence spoken twice never produces the same data, while text is identical no matter who types it, and one spoken word takes thousands of times more bytes than the same word in text. The model has to learn accents, pauses, breathing and timing as well as words and facts.

Block frames this as opening new research questions: how efficiently speech-only models learn language and facts compared with text models, how quickly they can pick up new languages, and what kinds of in-context learning they can do. It says Dessa knows a surprising amount for a model whose every fact came from raw sound.

Trained on one rack

Block says it worked out how to train Dessa on a single NVIDIA GB200 NVL72, a rack-scale system with 72 Blackwell GPUs. That is substantial hardware, but it is small compared with the clusters of tens of thousands of GPUs used to train frontier text models.

Block did not say how long training took, how many parameters the model has or how much audio it was trained on. Those details would be needed to judge how its efficiency claim compares with other speech models.

Why Block says it built Dessa

The team gives three reasons: curiosity, better understanding between humans and AI, and safety. On curiosity, it notes that babies learn to talk without ever seeing a written word and says it wanted to see whether a model could too. Its answer: it can, but it is very hard.

On safety, Block argues that AI keeps gaining power and independence while staying nearly deaf to the emotion in human voices. It asks readers to imagine being in serious pain while a voice AI has no idea and just keeps going, and calls Dessa a step toward fixing that problem.

Because a model that learned to talk from sound alone is new territory, Block says it wants outside eyes on it and is giving a small group of AI safety researchers early access to look for risks it has not anticipated. Applicants are asked to use a university, company or lab email address.

Who is behind it

Dessa comes from what Block calls a very small AI research lab inside the company, founded by former members of Dessa, the Toronto deep learning startup that Square acquired in 2020. The team says it has built deep learning models since 2012 and generative speech models for more than eight years, and that it reports directly to Jack Dorsey, Block's cofounder and chairman.

The project is led by Ragavan Thurairatnam, who signs the announcement. Block says more recordings will be shared on X.

What is missing

For now Dessa is a demonstration. There is no public model, no API, no pricing and no release date beyond a waitlist. Block has not said whether it will release weights, build Dessa into Cash App or Square, or keep it as a research project.

There are also no independent tests. The clips are chosen by Block, and the comparison with GPT-Live-1 is a single conversation rather than a controlled evaluation. A speech-only model may also be weaker at tasks that text models handle easily, such as precise facts, maths or following detailed instructions, and Block has not published results on those.

Why it matters

Voice is one of the busiest areas in AI this month. Mistral's Voxtral 3, ElevenLabs' realtime agent models and xAI's Grok Voice 3.0 all aim to make machine speech faster and more natural, but they lean on text-trained models underneath. Dessa is a bet on a different route: learning conversation directly from sound.

If Block's claims hold up under outside testing, a model that learns language from far less data, runs on a laptop and hears emotion natively would be a meaningful research result. It would also sharpen the safety questions Block itself raises, because a voice that laughs, gasps and sounds empathetic is more persuasive than one that obviously reads from a script. The early-access programme for safety researchers is the first place those questions will be tested.