Block presents Dessa, a full-duplex speech-to-speech model
Block, the company behind Square and Cash App, has presented Dessa, a speech-to-speech conversation model from its small AI research lab. Dessa's lead is Ragavan Thurairatnam. The page behind the announcement is billed with the line "the most authentic full-duplex speech model ever built"; that is the team's own framing, and no measure for it is given in the material available.
Full-duplex means the model listens and talks at the same time. The page says Dessa can backchannel (the small "mm-hm" acknowledgements people make while someone else speaks) and can cut in at natural moments in a conversation.
The developers say Dessa was trained from scratch only on human speech, with no text and no pretrained models. They claim it learned language from hundreds of times fewer words than large language models, though no exact figure is given. They also say it hears and produces emotion, laughter, sarcasm and pauses, which they argue text-based voice AI loses. Training used one NVIDIA GB200 NVL72 system. The developers say inference runs faster than real time and runs locally on a MacBook Pro M4 Max or better.
The team also makes a safety argument: voice AI that can't hear emotion is a safety problem. That framing fits the access plan. The team is offering early access to a small group of AI safety researchers through a form on the page.
One detail on the page concerns a demo conversation in which a third-party model, GPT-Live-1, appears. Block says it isn't affiliated with that model's developer.
No benchmark results, latency figures or numeric comparisons with other speech models are given in the available text. No release date, license, weights availability or pricing is stated.
Key facts
- Dessa is a speech-to-speech, full-duplex conversation model from Block's small AI research lab, led by Ragavan Thurairatnam.
- Its developers say it was trained from scratch only on human speech, with no text and no pretrained models, on one NVIDIA GB200 NVL72.
- The developers claim it learned language from hundreds of times fewer words than large language models, with no exact figure given.
- Inference is said to run faster than real time and locally on a MacBook Pro M4 Max or better.
- Early access is offered to a small group of AI safety researchers through a form on the page; no release date, license or pricing is stated.
Why it matters
Most voice assistants pass speech through text, and the developers argue that this loses emotion, laughter, sarcasm and pauses. Dessa is presented as a model that works on speech directly, from training onward, and that can listen and talk at once, backchannel and cut in. If the claims hold, that is a different design from text-first voice AI. The team also frames the gap as a safety issue: voice AI that can't hear emotion is a safety problem.
Who it affects
For now, mainly the small group of AI safety researchers who are being offered early access. Beyond them, the claims are of interest to teams building voice interfaces and to anyone following speech models trained without text. Block's other products, such as Square and Cash App, are not mentioned in connection with Dessa in the available text.
How to use it
There is no general release in the available material. The team is offering early access to a small group of AI safety researchers through a form on the page. The developers say inference runs locally on a MacBook Pro M4 Max or better. No release date, license, weights availability or pricing is stated.
How solid is it
Thin. Every capability above is a developer claim, and the available text is a summary of the page rather than the full article. No benchmarks, latency figures or comparisons with other speech models are given, the exact amount of training data is not stated, and nothing explains how the "most authentic" label is measured. The only text crawled from the site is the privacy notice for the dessa.com website, effective October 21, 2026, which says nothing about the model itself.
Risks and caveats
The headline claim is the team's own and is unverified. The "hundreds of times fewer words" comparison has no stated baseline or exact figure. Access is limited and no weights are mentioned, so outsiders cannot check the results. A demo conversation features a third-party model, GPT-Live-1, and Block says it isn't affiliated with that model's developer.