Block’s Dessa learned to talk only by listening, and it sounds human

Jack Dorsey’s company has shown a speech model trained from scratch on recordings of human conversation, with no text at all. It listens and talks at the same time, laughs mid-sentence, and runs on a high-end MacBook. It is not yet released.

By Roman Aglaganov26 OCT 2026North America

Illustration: two overlapping grey sound waves facing each other, one black and one grey, meeting in a single red peak — Illustration: AI Pulse.

Block, the company behind Square and Cash App, unveiled Dessa on 23 October, a speech model built by a small research lab that reports directly to Jack Dorsey. It is a full-duplex speech-to-speech model: sound goes in, sound comes out, and it can listen and speak at once, as people do.

No text at all

Most voice assistants turn your speech into text, have a language model write a reply, then read it aloud. Block argues that this throws away tone, timing, sarcasm and emotion. “Interacting with current voice AI can just plain suck,” it wrote. Dessa skips the text: Block says it was trained from scratch on nothing but recordings of human conversation, with no text and no pretrained model, using a single Nvidia GB200 NVL72 system.

In Block’s demo recordings, Dessa laughs in the middle of a sentence, gets louder when asked to speak up, says “yeah” while the other person is still talking, and finishes a sentence when someone forgets a film title. Block says it learned language from hundreds of times fewer words than a text model needs, runs faster than real time, and can run locally on a MacBook Pro with an M4 Max chip or better.

The caveats

There is no public access, technical paper, benchmark or price. Reviewers who listened to the samples found them uneven: some exchanges sound remarkably natural, others stiff. Block is instead giving a small group of AI safety researchers early access to look for risks it has not anticipated.

Block frames emotion as a safety issue: today’s voice AI cannot hear when someone is in distress. The same skill raises the obvious worry. A voice that sounds this human, running on a laptop, could make scam calls far more convincing.

Why it matters

Dessa is a test of whether models can learn language the way children do, from sound alone. If it works at scale, voice AI could stop sounding like a text chatbot reading aloud. Block, which already uses voice AI for ordering at Square, has an obvious commercial reason to want that, and a duty to decide carefully how such a voice is released.