Jack Dorsey's payments company Block has trained a new voice model, Dessa, entirely on human speech — using no text data and no pre-trained models, the company said in an official announcement. Block says the finished model runs locally on a MacBook Pro M4 Max or better, and that the small research lab behind the project trained it on a single rack of Nvidia servers. Running locally means a conversation is processed on the machine itself rather than streamed through a cloud service.

Block describes Dessa as the first AI it is aware of that can hold a lifelike live conversation after being trained exclusively on human speech. The compact research lab behind the model reports directly to Dorsey.

Full-duplex model held an unsteered demo with OpenAI's GPT-Live-1

Most voice assistants take what a person says, convert it into text, run the text through a language model, and then convert the result back into speech. Block argues that pipeline discards emotion, sarcasm, laughter, and what it calls "a sad voice that cracks."

Dessa instead turns speech directly into speech, with no text stage in between. Like two people in conversation, it talks and listens at the same time — audio flows both ways at once, rather than in the strict turns a transcribe-process-speak pipeline allows. The model knows when to cut in and when to chime in with a "yeah!" or "got it!", Block said, and it runs faster than real time.

According to Block, no one guides the model in the released recordings — Dessa chooses what to say, and when and how to say it. In the clips, it laughs mid-sentence, gasps at a surprise, and gets louder when asked to speak up. In one recording, Dessa held a conversation with OpenAI's GPT-Live-1. Block notes that the developer of GPT-Live-1 has not vetted or endorsed Dessa.

Announcing the model on X, project lead Ragavan Thurairatnam wrote:

"We made the first AI model you can have a realistic conversation with that was trained entirely on speech! Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs. This is a GPT3 moment for speech." pic.twitter.com/XVD8sSyJn8

— Ragavan Thurairatnam (@RagavanThurair) October 23, 2026

A GB200 NVL72 rack, and "hundreds of times fewer words"

Dessa was trained on a single Nvidia GB200 NVL72, a rack-scale system that links 72 of Nvidia's Blackwell GPUs to work as one machine. The lab says the model learned language from "hundreds of times fewer words" than a text-based LLM requires. Speech made the job tougher, Block explained: a spoken word carries thousands of times more bytes than the same word in print.

The lab was founded by former members of Dessa, a Toronto deep learning startup, and its team has been building generative speech models for more than eight years. Thurairatnam heads the project and called the release "a GPT3 moment for speech" on X — a nod to GPT-3, OpenAI's 2020 text model that took machine-generated writing from research demo to widely used tool, and his shorthand for placing speech at a similar turning point.

Dessa is not publicly available. Block is running a waitlist and granting early access to a small group of AI safety researchers to surface risks the team has not foreseen. The announcement gave no timeline for wider availability.

Three things motivated Block to build the model. Two are curiosity and better understanding between humans and AI. The third is safety: Block points to the scenario of a person who is in a of pain while a voice AI fails to notice and keeps talking — precisely the kind of emotional signal that a text-first pipeline can strip out before the model ever hears it.

The announcement comes during a period of change at Block. As Cryptopolitan reported, the company (NYSE: XYZ), formerly Square, laid off 40% of its staff in February. In July, Block launched Buzz, a workplace chat app where humans and AI agents can share channels.