Jack Dorsey's Block on October 23, 2026 unveiled Dessa, a speech-to-speech artificial intelligence (AI) model. "Speech-to-speech" means the model takes a voice request and answers in voice too, with no text involved in the process at all. According to the company, Dessa was trained from scratch exclusively on human conversation — books, web pages or other written texts were not used in training at all. This approach sets it fundamentally apart from nearly all large AI models of today, which are trained mostly on giant text corpora from the internet.
According to The Verge, the announcement came with a public waitlist: anyone interested can sign up to try the model. At the same time, the company is giving early access to a small group of researchers working on AI safety — the team believes such a model opens new territory for research and may carry risks that have not yet been anticipated.
Full-Duplex: A Model That Talks While Listening
Dessa's key technical differentiator is that it runs in full-duplex mode (listening and speaking at the same time). Audio goes in, the system produces audio, and speech is never converted into written words along the way. In practice, this means the model keeps listening even while it speaks: if the interlocutor interrupts it mid-sentence, pauses, laughs or changes tone — it senses all of this in real time and responds naturally. This is behavior close to how people converse: we also talk while listening, interrupt our interlocutor, and adjust our replies based on their tone.
According to the Dessa team, this is exactly the problem with today's voice AI systems: they first convert human speech into text, and in that process a huge amount of data carried by tone, pauses and emotion is lost. As a result, even the most powerful models cannot hear their interlocutor's mood. The team explains the creation of the model with three reasons: curiosity (or, in their own words, delight), better mutual understanding between humans and AI, and safety.
"We trained Dessa from scratch on human conversation alone. Sounds! No text, no pretrained models." — Block, official Dessa page
The Model That Never Read a Book — It Only Listened
The second feature that sets Dessa apart from other models is the nature of its training data. Typical large language models are trained on billions of words from books, web pages and other written sources. Dessa, by contrast, learned only by hearing the sounds of live human conversation — with no written text and no pretrained models. In other words, it learned language not from books but by listening to people — just as babies learn to speak without ever seeing writing.
This approach directly shapes the model's behavior. According to the company, when Dessa hears something funny, it laughs mid-sentence like people do — while other models wait for the interlocutor to finish and then laugh "on command." If asked to speak louder, it raises its voice — none of this was taught to it specifically; it picked it all up from the conversations it heard. The company describes Dessa as "the world's first AI that learns from human speech and can hold realistic live conversation."
Runs Locally on a MacBook
According to Block, Dessa can run locally on a MacBook Pro M4 Max or a more powerful computer. This is an important practical detail: the model does not need to call cloud servers for every request and can run on the user's own device — meaning voice data may never leave the device.
Access to the model is limited for now. The announcement came with a public waitlist — anyone interested can sign up. In parallel, the company is giving early access to a small group of researchers in AI safety. The reason was stated openly: a model trained only on sound is new territory for research, and the team wants outside specialists to find risks it could not foresee itself. Safety is cited as one of the three main reasons the team built the model.
Project Lead and the History of the Dessa Name
The project is led by Ragavan Thurairatnam — head of Block's internal AI research lab. Thurairatnam co-founded the original Dessa startup, which Block acquired in 2020. The new product's name revives that brand: the name of the startup acquired six years ago now returns as a project with a fundamentally new approach.
So Dessa is a voice AI trained without written texts — only on the human voice and live conversation. It talks while listening, senses interruptions and laughter in real time, and can run locally on a personal computer. For now, the model is being offered through the waitlist and as early access for safety researchers.
Note that Block's Square division is also working on voice AI — we previously wrote about the company introducing an AI-powered voice ordering kiosk.
