Block's in-house AI research lab has unveiled Dessa, a full-duplex speech-to-speech model that the company says was trained from scratch on nothing but recorded human conversation, with no text and no pretrained models. Block says Dessa runs faster than real time and can run locally on a MacBook Pro with an M4 Max chip or better.
Full-duplex means Dessa listens and talks at the same time. Audio goes in and audio comes out, with no transcription step in between. Block says the model learned from real conversations when to interject and when to backchannel with short responses such as "yeah" or "got it," and that it can hear and produce emotion, sarcasm, laughter, sighs and voice cracks.
Block says it trained Dessa on a single Nvidia GB200 NVL72 system, and that the model learned language from "hundreds of times fewer words" than large language models. The company has not published a parameter count, dataset size, latency figures or formal benchmark results for the model.
Key facts
- 01Block's AI research lab unveiled Dessa, a full-duplex speech-to-speech model, as a research preview on October 23, 2026.
- 02Block says Dessa was trained from scratch on human conversation only, with no text and no pretrained models.
- 03Training ran on a single Nvidia GB200 NVL72 system, and the model runs locally on a MacBook Pro M4 Max or better.
- 04Early access is limited to a small group of AI safety researchers; a public waitlist is open.
- 05The lab is led by Ragavan Thurairatnam and reports directly to Block cofounder and chairman Jack Dorsey.
Block's case for the approach is that most current voice assistants convert speech to text before reasoning about it, which strips out emotional cues. The company gives the example of an AI that cannot tell when a person is in serious pain. Block also argues that speech is harder to model than text because spoken data never repeats exactly and a spoken word takes thousands of times more bytes than its written form.
The team says it took its cue from the way babies learn to talk without ever seeing written words. Block's announcement opens with the line "You learned to talk by listening," and calls Dessa "the most authentic full-duplex speech model ever built."
In place of benchmarks, Block released recordings it describes as live and unsteered, with subtitles added afterward. In them, Dessa laughs mid-sentence, raises its volume when asked, finishes a speaker's sentence when they forget a movie title and gasps at a surprise anniversary announcement. One clip pairs Dessa with GPT-Live-1, a third-party voice model; Block notes that model's developer is not affiliated with Block and has not reviewed or endorsed Dessa.
Dessa is a research preview, not a product. Early access is limited to a small group of AI safety researchers, who must apply with a name and an institutional email from a university, company or lab. Block says it is giving that group access "to find the risks we haven't anticipated."
“A model that learned to talk from sound alone is new territory, and we want outside eyes on it.”— Ragavan Thurairatnam, Dessa lead at Block
A public waitlist is open for everyone else, and Block says launch links are "coming soon." The company also says more recordings will be posted on X.
The lab behind Dessa is a small team founded by former members of Dessa, the Toronto deep learning startup Block acquired in 2020. The team says it has built deep learning models since 2012 and generative speech models for more than eight years, and it reports directly to Block cofounder and chairman Jack Dorsey.
Ragavan Thurairatnam, who leads the lab and co-founded the original Dessa startup, announced the model on X and called it "a GPT3 moment for speech." Block lists six researchers on the project: Louay Hazami, Hashiam Kadhim, Qi (Alex) Liu, Rayhane Mama, Aydin Polat and Marc Tyndel.
Block also lists research questions it has not yet answered, including how efficiently speech-only models learn language compared with text-trained models, how well they pick up new languages and what kinds of in-context learning they can perform.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
