Block’s Dessa Voice Model Learned Only From Speech

Block, Jack Dorsey’s payments company, has shown off Dessa, a speech-to-speech model that its research lab trained from scratch on human speech alone. There is no text in its training data and no pretrained language model inside it, and Block says it is probably the first realistic live conversational AI built that way. Dorsey announced it on 24 October.

Dessa is full-duplex, meaning it listens and talks at once rather than waiting for its turn. On Block’s demo page it backchannels while the other person talks, laughs in the middle of sentences, and is meant to pick up and produce emotion, sarcasm and pauses. Block’s argument, as it explained to TechCrunch, is that most voice assistants sound flat because text drives the speech they generate, so emotion has to be bolted on afterwards. Dessa’s knowledge comes from the conversations it was trained on rather than a separate LLM, which also means it has seen what Block calls hundreds of times fewer words than a language model. The company says the roughly 20 demo recordings are unsteered, with subtitles added afterwards, and two of them feature OpenAI’s GPT-Live-1, a model Block stresses it has no affiliation with.

The compute numbers are small by frontier standards. Block says Dessa was trained on one Nvidia GB200 NVL72 rack and runs faster than real time on a laptop, specifically a MacBook Pro with an M4 Max chip or better. That local footprint matters for the business case: Square already offers sellers AI voice ordering, and an on-device model would cut the cost of putting a voice on more of Block’s products.

The lab behind it was founded by former members of Dessa, the Toronto deep-learning startup whose name the model takes, and it reports directly to Dorsey; Ragavan Thurairatnam leads it. Nobody outside Block can try the model yet. There is a waitlist, and early access is going first to a small group of AI safety researchers with institutional email addresses, whom Block wants to test it for ways a voice agent could deceive people.

Full-duplex voice is now crowded, with Microsoft’s MAI-Realtime and Thinking Machines’ TML-Interaction-Small among the models chasing the same natural back-and-forth. What sets Dessa apart is what it leaves out. A model that learned language only by listening may sound more human than anything built on text, but the open question for a company that processes payments is whether it can be as dependable about an order total as it is about a laugh.