Jack Dorsey's Block Built a Voice AI From Speech Alone, and It Runs on a Laptop.

Jack Dorsey's payments company Block released a voice model on Friday, October 23, 2026, and the thing that makes it strange is what is missing from it. Dessa was trained from scratch on recordings of human conversation — no text, no pretrained language model underneath. Block's own announcement puts it plainly: "We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models" 1.

Most voice assistants run on a pipeline: your speech becomes text, a language model writes a reply, and a speech model reads that reply out loud. Block's argument is that the middle step throws away the part of speech that carries meaning — tone, timing, a sigh, a voice cracking — and that this is why talking to them feels slow and stilted 1. Dessa skips it: audio in, audio out. Dorsey called it "the most authentic full-duplex speech model ever built," with full-duplex meaning it listens and speaks at the same time, so it can interrupt, backchannel with a "yeah" or a "got it", and laugh in the middle of a sentence instead of waiting its turn 2.

The demos Block published show what that buys. The model laughs the moment something is funny, gets louder when a person asks it to, finishes a movie title the speaker has blanked on, and answers a piece of news with what the company describes as teary gratitude 1. It also mistakes one speaker for somebody named Ronaldo, which The Verge noted gives the whole thing the feel of "an awkward new friend" 3.

Learning language from sound is the hard way round. Block points out that a spoken word takes thousands of times more bytes than the same word typed, and that speech never repeats itself, so the same sentence recorded twice is never the same data 1. The company says it trained the model on a single Nvidia GB200 NVL72 rack and that it picked up language from hundreds of times fewer words than a text model needs, and that inference now runs faster than real time on a MacBook Pro with an M4 Max chip or better 14.

Block is a payments company — Square, Cash App, the Bitkey wallet, the Proto mining rig — and the lab behind Dessa was founded by former members of the Toronto deep learning startup Dessa, which Square acquired in 2020. It reports directly to Dorsey 15. Square already uses voice AI to take restaurant orders, and a model that sounds this good running on-device points at cheaper versions of the same thing 4.

Dessa is a research preview with no public access. Block has not published the model's size, its benchmark results or a release date, and every audio clip it has shared is a company demonstration rather than an independent test 5. What Block has opened is a call for a small group of AI safety researchers to take early access and find the risks it has not anticipated 1. Ragavan Thurairatnam, who leads the team, called the release "a GPT3 moment for speech" 4.

The reaction split the way you would expect a human-sounding voice to split it. Filmmaker Justine Bateman asked why an AI voice needs to sound human at all, and answered her own question: "So you can trick people into lowering their guard." Journalist Shiv Aroor put the practical version: "Scam calls about to attain pandemic level." Others wanted it for the good cases — Binance founder Changpeng Zhao said it "might be very good for GiggleAcademy", his education project, because "teaching kids in current 'AI voice' was a challenge" 5.

Block's stated reason for building it is partly curiosity and partly this: "AI keeps gaining power and independence, and it's still nearly deaf to the emotion in our voices. Imagine being in serious pain while a voice AI has no idea and just keeps going. That's the safety problem" 1. A model that hears the ache is the fix they are proposing. It is also, by construction, a model that can perform an ache it does not have.

"SOUND ONLY" is an original Miami bass song and music video built from that reporting — one song, one day's news, with the text step pulled out of the pipeline as the hook.