Block builds Dessa voice AI from human audio alone on a single Nvidia rack

Jack Dorsey’s Block trained its new voice model, Dessa, on human speech only, not on any text or any pre-trained models. Block says that the finished model can be run locally on a MacBook Pro M4 Max or better. The lab trained it on a single rack of Nvidia servers.

Block says Dessa is the first AI it knows of that can carry on a lifelike live conversation and has been trained only on human speech. The tiny research lab behind it reports directly to Dorsey. Full-duplex model talks with OpenAI’s GPT-Live-1 in unsteered demos A lot of voice assistants take what people say, turn it into text, run it through a language model, and then turn the text back into speech.

As Block puts it, the pipeline throws out emotion, sarcasm, laughter, and a sad voice that cracks. Dessa is a model that turns speech into speech, with no text in between. Like two people, it talks and hears at the same time.

It knows when to cut in and when to chime in with a “yeah! ” or “got it! ” Block said.

It runs faster than real time. Block says no one guides the AI model in the recordings. It chooses what to say, when and how.

In the recordings, Dessa laughs mid-sentence, gasps at a surprise and gets louder when asked to speak up. In one of the recordings, Dessa spoke to OpenAI’s GPT-Live-1. Block says the developer of GPT-Live-1 has not vetted or endorsed Dessa.

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.

— Ragavan Thurairatnam (@RagavanThurair) October 23, 2026 A GB200 NVL72 taught the model language from hundreds of times fewer words Dessa trained on a single Nvidia GB200 NVL72, a rack-scale system. The lab says the model learned language from “hundreds of times fewer words” than a text-based LLM needs. Speech made the job tougher, said Block.

A spoken word is thousands of times more bytes than the same word in print. The lab was founded by former members of Toronto deep learning startup Dessa. The team has been building generative speech models for over eight years.

Ragavan Thurairatnam heads the project, and he called the release “a GPT3 moment for speech” on X. Dessa is not publicly available. Block is running a waitlist and giving early access to a small group of AI safety researchers to surface risks the team hasn’t foreseen.

Three things led Block to build the model. Two are curiosity and better understanding between humans and AI. The third reason?

Safety. Block imagines a person who is in a lot of pain, but a voice AI doesn’t notice and keeps talking. Cryptopolitan reported that in February, Block (NYSE: XYZ) laid off 40% of its staff.

Block Inc. put out Buzz in July. It’s a chat app for the workplace where humans and AI agents can share channels.

The smartest crypto minds already read our newsletter. Want in? Join them.