Block, the company behind Square and Cash App, has introduced Dessa, a voice model that learned to hold a conversation from recorded human speech and nothing else. The team says no text and no pretrained models went into it. It also says Dessa is, “as far as we know,” the first realistic live-conversation AI trained only on speech, a claim nobody outside the lab has checked yet.
Dessa comes from a very small research lab at Block led by Ragavan Thurairatnam, which reports directly to Jack Dorsey, the company’s cofounder and chairman. According to the announcement post on Block’s site, its founders came out of the Toronto startup that was also called Dessa. The team says its deep learning work dates to 2012 and that it has spent over eight years on generative speech.
Most voice assistants work by transcribing what you say, handing the text to a language model, and reading the answer back. Block’s post argues that this throws away how you sound: sarcasm, a sigh, a cracking voice. Dessa skips the transcript. It takes audio in and puts audio out, and it is full-duplex, so it listens while it talks and decides on its own when to cut in or drop a quick “got it.”
The deployment detail is the one builders will notice. Block says the model generates speech quicker than the clock, and that it can run on the user’s own machine, specifically a MacBook Pro with an M4 Max chip or better. A model that fits on a high-end laptop can in principle hold a voice conversation without a server on the other end, though the post does not say how much memory it needs or how many parameters it has.
The training numbers are the boldest part of the pitch. Block says the language Dessa absorbed amounts to hundreds of times fewer words than text-trained models are fed, and that it was trained on a single NVIDIA GB200 NVL72 system. The lab is candid about why that is surprising: one spoken word occupies thousands of times the storage of that word typed out, and speech never repeats exactly, even when the same person says the same sentence twice. The post gives no hours of audio, no benchmark scores, and no comparison against a text-trained model.
What the post does offer is recordings. In the clips, Dessa laughs mid-sentence, finishes a person’s sentence when they blank on a movie title, and gasps at an anniversary surprise. In one demo it chats with GPT-Live-1, a model from another developer, and the post says that developer has no tie to Block and has neither reviewed nor endorsed Dessa. These are clips the team chose to publish, and Block says more will follow on X.
Block lists three reasons for building it: curiosity, better understanding between people and AI, and safety. The safety case rests on a gap. The post says AI systems are gaining power and autonomy while remaining close to deaf to how people feel when they speak. “Imagine being in serious pain while a voice AI has no idea and just keeps going,” Thurairatnam writes, calling Dessa a step toward fixing that.
Block is also asking for scrutiny. It plans early access for a handful of outside AI safety researchers, who are meant to find risks the team has not anticipated. For everyone else there is a waitlist, and the launch links on the page read “coming soon.”
The post names one open question as the next test: how efficiently speech-only models pick up new languages. If the hundreds-of-times data claim holds up there, it would matter most for languages with little written text online, and independent results from the safety group are the first place that evidence could appear.
Reported by Block in its announcement post “Dessa: the conversation model trained only on speech,” by Ragavan Thurairatnam; the post carries no publication date.
