Block's Dessa speech model is a voice AI that learned to talk without ever reading a word. Block, the company behind Square and Cash App, has released Dessa as a research preview: a full-duplex model that takes audio in and sends audio out, with no text step in between. Its research team says it was trained from scratch on human speech and sounds only.
Most voice assistants today still work in a chain. Your speech is turned into text, a language model writes a reply, and a second system reads that reply aloud. Dessa skips the text completely. That design choice is what makes it interesting, and it is also why Block is only letting a small group of safety researchers try it for now.
What the Dessa speech model does
According to Block's Dessa page, the model listens and speaks at the same time. In practice that means it can:
- interrupt you, or be interrupted, the way people do in real conversation;
- add small "backchannel" sounds while you talk, such as "yeah!" or "got it!";
- hear and produce emotion, sarcasm, laughter and sighs;
- recall facts and finish a speaker's sentence.
Block says inference runs faster than real time, so there is no awkward pause before it answers. The page gives no latency figures, model size or benchmark scores. It does say the model runs locally on a MacBook Pro with an M4 Max chip or better.
The recordings on the page are live conversations in which Dessa picks its own words, timing and voice. Block notes that subtitles were added afterwards. One demo pairs Dessa with a third-party AI, GPT-Live-1, and the page states that its developer is not affiliated with Block and has not reviewed or endorsed Dessa.
Trained on sound, not text
The training approach is the main claim. Block's team says Dessa was trained from scratch on human speech and sounds, with no text and no pretrained models, on a single NVIDIA GB200 NVL72 system. The team adds that, as far as it knows, it is the first AI of its kind trained only on human speech that can hold a realistic live conversation.
The page also says Dessa learned language from "hundreds of times fewer words" than large language models use. That is a comparison, not a benchmark, and Block publishes no test results to back up how well it reasons or how accurate its facts are. Readers should treat it as an early research result.
Who built it
Dessa comes from a small AI research lab inside Block, founded by former members of Dessa, a Toronto deep learning startup. Block says the lab 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 co-founder and chairman. The project is led by Ragavan Thurairatnam.
Who can use Dessa right now
Not most people. The page lists "launch" links as coming soon and offers early access to a small group of AI safety researchers, who are asked to contact the team by email. There is no public waitlist and no pricing.
Thurairatnam explains the caution on the page: "A model that learned to talk from sound alone is new territory, and we want outside eyes on it." Block says it wants safety researchers to find risks the team has not anticipated.
Why it matters for users and builders
Voice is where a lot of AI use is heading, from customer service lines to in-car assistants. Today's systems lose information when they turn speech into text: tone, timing, hesitation, a laugh. Block's argument is that a model which keeps all of that could understand people better, including someone who is upset or in pain.
There is also a practical angle for regions where many people prefer speaking to typing, including much of Africa. A model that learns from speech rather than text could, in principle, be trained for languages that have little written data online. Block has not said anything about language coverage, so that remains a question rather than a promise.
The same strengths bring risks. A voice that sounds natural, reacts to emotion and can run offline on a laptop raises familiar worries about impersonation, scams and emotional dependence. Those are the kinds of issues Block says it wants outside researchers to probe before any wider release.
How Dessa compares with other launches this week
It has been a busy week for model releases. Moonshot's Kimi K3.1 Pro arrived with a 1M-token context, as we explained in our Kimi K3.1 Pro guide, and open-source text-to-speech models such as Gradium's small Phonon are pushing voice onto phones and laptops. Dessa stands apart because it is not a text model with a voice bolted on. Whether that approach can match text-trained models on knowledge and reasoning is the open question.
What to watch
- Any published benchmarks or a technical paper from Block's team.
- Details on languages, accents and model size.
- When the "coming soon" launch links go live, and whether a public demo follows the safety-researcher phase.
Until then, Dessa is best read as a research signal: a serious company betting that the next step in voice AI is to stop turning speech into text at all.
