History
The History of Dessa: Toronto Startup, Joe Rogan Deepfake, Block Lab
Before Dessa was a Block research preview announced on October 23, 2026, Dessa was a small Toronto startup with a name the AI world first heard for a very different reason: a fake Joe Rogan interview. The history matters, because the team, the city and even the name of Block's new voice model all come from that earlier company — and the path from deepfake notoriety to a safety-first voice lab explains a lot about how the 2026 launch was handled.
2016: DeepLearni.ng in Toronto
According to Block's own account and the editorial bible for this site, Dessa began in 2016 as DeepLearni.ng, a Toronto deep-learning startup that later took the name Dessa. Toronto at that time was one of the world's most important machine-learning hubs, shaped by university research and a dense community of applied AI companies. Dessa belonged to that scene: a small, research-heavy team working on practical machine learning, not a consumer chatbot company chasing headlines.
That origin is important for understanding the 2026 model. The present-day Dessa lab inside Block is described by Block as founded by former members of that Toronto startup. In other words, this is not a team Block assembled from scratch for a press release. It is a surviving research group with a shared history that predates the acquisition by years, and Block says the wider team has built generative speech models for more than eight years.
2019: the Joe Rogan deepfake
Dessa became famous in 2019 for a demonstration that now reads like a warning from the past: a synthetic Joe Rogan voice, convincing enough to show how easily recorded speech could be imitated. At the time it was presented as research and as a public demonstration of what voice cloning could do — and why society should think about detection and misuse before the technology became ordinary.
We summarise that episode carefully on this site because Block's verified-facts record confirms the outline — founded in Toronto, famous for the 2019 Joe Rogan deepfake — without turning folklore into detail Block has not published. The significance is not the stunt itself. It is that the people who would later build a human-sounding conversational model had already seen, early, how quickly a realistic voice can become a trust problem. That experience helps explain why, in 2026, Block put AI safety researchers first in line for access rather than opening Dessa to everyone at once, a rollout we examined in is Dessa AI safe?
February 7, 2020: Square acquires Dessa
On February 7, 2020, Square — the company now called Block — acquired Dessa. Acquisitions of research startups often dissolve the original team into unrelated product work. Here, the through-line is unusually direct: the acquired Toronto group became the foundation for a Block research lab that kept the Dessa name, kept working on machine learning and speech, and eventually produced a model that carries the same name.
Block itself has changed names and shape in the meantime. Square became Block, Inc., the parent of Square, Cash App, Bitkey and Proto, led by co-founder Jack Dorsey. That parentage matters because a payments company lives on trust: moving money, verifying identity and preventing fraud are its daily business. A voice model from such a company was never going to be treated as a toy demo, however playful its launch recordings sound.
2020 to 2026: the quiet years
Block has not published a detailed public diary of what the Dessa team did between acquisition and announcement, and this article will not invent one. What is verified is the endpoint: a tiny lab inside Block, led by Ragavan Thurairatnam, with six researchers and three interns named on Block's official page, reporting directly to Dorsey. Thurairatnam has called the launch "a GPT-3 moment for speech" — a claim about a turning point, not a benchmark, and one independent testing will have to judge.
The result they announced was unlike the startup's earlier voice work in one crucial way. The Rogan demonstration imitated a specific person's recorded voice. Dessa, by contrast, is presented as a conversational model in its own right: trained from scratch on human conversation audio only, with no text and no pretrained models, learning language the way Block argues babies do — by listening. For the technical story, see how Dessa AI works.
October 23, 2026: Dessa becomes a voice model
When Block unveiled Dessa, it stressed live, unsteered recordings: the model laughing mid-sentence, backchannelling with "yeah" and "got it", finishing a speaker's sentence, gasping with what sounds like tearful gratitude, and even talking with OpenAI's GPT-Live-1 — with Block noting OpenAI is unaffiliated and has not endorsed it. Reviewers found the demos mostly natural, with occasional awkwardness, including a moment where Dessa called a speaker "Ronaldo".
Block also stressed efficiency: training on a single NVIDIA GB200 NVL72 system, learning from hundreds of times fewer words than text language models, and inference faster than real time locally on a MacBook Pro with an M4 Max chip or better. Those are Block's claims, not independent measurements, and Block published no parameter count, benchmarks or training hours.
Why the history shapes what happens next
A team that became famous for a deepfake, was acquired by a payments company, and then spent years inside that company building speech models is unlikely to be naive about impersonation. Block's chosen order — research preview, waitlist at block.xyz/dessa, safety researchers first — fits that history. It also sets expectations: if early access reports focus on misuse, disclosure and safeguards before they focus on charm, that will be continuity, not a change of direction.
Dessa is still not a public product. Block has announced no pricing, no API, no weights and no release date. But its name now carries two meanings at once: a Toronto startup that warned the world about fake voices in 2019, and a 2026 model asking whether a machine that learns only by listening can sound more human than one that learns by reading. The next chapter will be written by the safety researchers who get access first — and we will cover their findings here.
Sources: Block's official Dessa page (block.xyz/dessa) for the team, acquisition lineage and 2026 model claims; DessaAI Daily's verified-facts record summarising Block, The Verge and TechCrunch coverage, October 2026.
Frequently asked questions
When was Dessa founded and where?
Dessa began in Toronto in 2016 as DeepLearni.ng, a deep-learning startup that later took the name Dessa. Block's 2026 voice lab was founded by former members of that Toronto company.
What was Dessa's 2019 Joe Rogan deepfake?
In 2019 Dessa became famous for a synthetic Joe Rogan voice demonstration that showed how convincingly recorded speech could be imitated, raising early questions about voice cloning and misuse.
When did Square acquire Dessa?
Square, now called Block, Inc., acquired Dessa on February 7, 2020. The Toronto team became the foundation for the Block research lab that later announced the Dessa voice model.
🚀 Get Dessa launch alerts
Dessa is still a research preview. Join the free DessaAI Daily list and we will email you the day access opens — plus the week's biggest Dessa stories, once a week. No spam, no hype.
Form not loading? Subscribe on the signup page →