Dessa, announced October 23 by Block's in-house AI research lab, is a full-duplex speech-to-speech model. According to Block, it was trained from scratch on nothing but human conversation. No text. No pretrained models. It learned to hold a real-time conversation by listening.

The claim is substantial. Block says Dessa is the world's first AI that can carry on a realistic live conversation after being trained only from human speech. Most voice AI systems work differently: text goes in, text gets processed by a large language model, and then the output is synthesized into speech. Dessa skips the text entirely. Audio in, audio out.

Full-Duplex Matters

Full-duplex means the model listens and talks at the same time. Ragavan Thurairatnam, who leads the lab and co-founded the original Dessa startup that Block acquired in 2020, described it as "a GPT3 moment for speech."

The distinction sounds technical but shows up immediately in demos. When something is funny, Dessa laughs mid-sentence, the way a human would, rather than waiting for the speaker to finish before inserting a canned response. It backchannels. It cuts in. When asked to speak louder, it does. Block claims nobody taught it that behavior explicitly.

Sample recordings on the Dessa page show conversations about Christopher Nolan films, a surprise trip to Paris, and an exchange where Dessa finishes someone's sentence when they blank on a movie title. Block also released a clip of Dessa talking with OpenAI's GPT-Live-1 voice model. In that exchange, Dessa reportedly comes across as more natural.

The conversations are not flawless. In one widely circulated demo, Dessa opens by addressing the speaker as "Ronaldo." The speaker, presumably, was not Ronaldo. Block chose to show its work, errors included.

A Small Team, A Large Technical Bet

The lab behind Dessa is small. Block lists six researchers: Louay Hazami, Hashiam Kadhim, Qi (Alex) Liu, Rayhane Mama, Aydin Polat, and Marc Tyndel, plus several interns. The team reports directly to Jack Dorsey. Thurairatnam co-founded the original Dessa, a Toronto deep learning startup known for deepfake detection work, which Block acquired in 2020. Several members of that team now staff the lab.

Advertisement

Training required a single NVIDIA GB200 NVL72 system, Block says. The company claims Dessa learned language from hundreds of times fewer words than typical large language models, though it also had to absorb far more than language: accents, pauses, breathing, sarcasm, humor, inflection, timing, when to interrupt. Block's page includes an example sentence that changes meaning depending on which word is stressed. Text encodes none of that.

Safety as Justification

Block gives three reasons for building Dessa: curiosity, better understanding between humans and AI, and safety. The safety argument is notable: AI systems keep gaining power and independence while remaining largely deaf to emotional signals in voices. Block's page poses the scenario of someone in serious pain while a voice assistant barrels ahead, unaware. Dessa, the company argues, is a step toward fixing that gap.

Block is offering early access to a small group of AI safety researchers to find risks the team has not anticipated. A public waitlist is open. The model can run locally on a MacBook Pro M4 Max or better, which is unusual for models of this type. Inference runs faster than real time, according to Block, so there is no awkward pause before responses.

Open research questions remain: how efficiently speech-only models learn language and facts compared with text-trained models, how well they pick up new languages, and what kinds of in-context learning they can perform. Dessa is a research preview. Launch links are listed as "coming soon."