Block has unveiled Dessa, a speech-to-speech AI model trained from scratch on human conversation alone — no text, no pretrained models — that can hold a live, two-way spoken conversation. Announced on October 23 by the payments company’s small in-house research lab, Dessa is a research release for now: early access goes to a small group of AI safety researchers, with a public waitlist open.
October 23, 2026 — Why it matters in one line: almost every voice assistant today turns speech into text and back again, throwing away tone and timing; Dessa is a bet that skipping text entirely produces a more natural — and safer — voice AI.
What Block built
Dessa is what Block calls a true speech-to-speech model: audio goes in, audio comes out, with no written words in between. It runs in full-duplex mode, meaning it can listen and talk at the same time, the way people do on a phone call rather than taking walkie-talkie turns. Block says it learned from real conversation when to cut in, when to backchannel, and how emotion sounds — sarcasm, laughter, a frustrated sigh, a voice cracking with sadness. Inference runs faster than real time, and the model runs locally on a MacBook Pro with an M4 Max chip or better.
The training approach is the headline. Block says Dessa learned language purely by listening, from hundreds of times fewer words than large language models typically train on, and that the team worked out how to train it on a single NVIDIA GB200 NVL72 system. The company frames it — in its own words, “as far as we know” — as the first AI that can hold a realistic live conversation after being trained only from human speech. That is a company claim, and Block has published no independent benchmark to back it up.
What the demos show — and their limits
Block published live recordings of people talking with Dessa and says nobody steers the model during them: it chooses what to say, when, and in what voice. In the clips, Dessa laughs mid-sentence when something is funny, finishes a person’s sentence when they blank on a movie title, and holds a conversation with GPT-Live-1, a third-party voice model. Coverage also noted rough edges — in one exchange the model mistakes the speaker for someone named “Ronaldo” — worth remembering: these are curated recordings, not a public product you can test today.
The team and the reasoning
Dessa comes from a small Block research lab led by Ragavan Thurairatnam and founded by former members of Dessa, the Toronto deep learning startup Block acquired in 2020; the lab reports directly to Block cofounder and head Jack Dorsey. The team gives three reasons for building it: curiosity, better understanding between humans and AI, and safety. Their safety argument is specific — voice AI that flattens speech into text cannot hear that a speaker is in pain or distress, and keeps going regardless. A model that hears emotion natively, they argue, is a step toward fixing that.
Why it matters
Treat Dessa as a research signal, not a product launch. There is no general availability, no pricing, and no published evaluation — access is limited to safety researchers, plus a waitlist with launch links promised soon. But the direction matters for anyone choosing voice tools today: the current generation of voice agents is built on text pipelines, and Block has shown a credible alternative path where timing, interruption and emotion are the model’s native language. If speech-native models hold up outside curated demos, expect the “sounds human” bar for voice AI to move quickly.
Frequently asked questions
Can I try Dessa now?
Not generally. Block is giving early access to a small group of AI safety researchers, and a public waitlist is open on the Dessa page. The company says launch links are coming soon but has given no date.
How is Dessa different from other voice assistants?
Most voice AI converts your speech to text, thinks in text, then speaks a reply. Dessa was trained on sound alone and works speech-to-speech, so it can react to tone, pauses and interruptions while you are still talking.
What hardware does Dessa need?
Block says it runs locally on a MacBook Pro with an M4 Max chip or better, and that inference runs faster than real time.
Is Dessa better than ChatGPT’s voice mode?
There is no independent evidence either way yet. Block’s demos include a conversation between Dessa and OpenAI’s GPT-Live-1 that Block presents favourably, but no head-to-head benchmark has been published.
Sources: Block (Dessa page, block.xyz), The Verge, Unite.AI, Crypto Briefing.