Block, the payments company co-founded by Jack Dorsey, has introduced Dessa, an AI voice model built on an unusual premise: everything it knows was learned from recordings of people talking, with no written text and no pretrained models involved. The model was unveiled on October 23 through a page on Block's website and posts on X, and it is not yet a product. Block is offering early access only to a small group of AI safety researchers, whom it invites to look for risks the team has not anticipated, alongside a public waitlist.
Dessa is a speech-to-speech system. Audio goes in and audio comes out, skipping the usual pipeline in which speech is transcribed, a language model reasons over the text and a synthesiser voices the answer. It is also full-duplex, so it listens while it speaks. Block says that lets the model cut in, offer short acknowledgements such as "yeah" or "got it", laugh mid-sentence and pick up on sarcasm, sighs or a breaking voice, cues that a transcript throws away. The company adds that inference runs faster than real time and that the model runs locally on a MacBook Pro with an M4 Max chip or something more powerful.
The team is small. Block describes it as a research lab started by people who previously worked at Dessa, a Toronto startup focused on deep learning whose name the model now carries, and says it reports directly to Dorsey. Block bought that startup in 2020, and the lab is led by Ragavan Thurairatnam, one of its co-founders, according to Crypto Briefing. Block's page lists six researchers plus a handful of interns and says the group has built generative speech models for more than eight years.
Block makes training claims it has not yet backed with a paper. It says Dessa picked up language from a word count hundreds of times smaller than what large language models consume, and that the whole model was trained on a single NVIDIA GB200 NVL72 rack-scale system. The company argues that speech-only learning is far harder than text: no sentence is ever spoken the same way twice, and one word said aloud needs thousands of times as many bytes as the same word typed.
So far the public evidence consists of curated recordings. Block posted live clips in which Dessa supplies a film title a speaker cannot remember, reacts with what sounds like tearful thanks to news of a surprise trip, and chats with GPT-Live-1, a voice model The Verge identifies as OpenAI's; Block stresses that the other model's developer is not affiliated with it and has not reviewed Dessa. The Verge, which listened to the samples, found the exchanges mostly natural but occasionally awkward and uncanny, citing a clip in which the model greets its human partner as someone called Ronaldo.
Block gives three motives for the work: curiosity, better mutual understanding between people and AI, and safety. Its safety case is that increasingly capable AI systems remain almost deaf to emotion in the human voice, so a model that hears distress directly might respond more appropriately to someone in pain. The flip side is obvious, as Crypto Briefing noted: a voice that sounds convincingly human and reads feelings also raises questions about impersonation and trust, which may explain the restricted rollout.
Much remains unknown. Block has not published benchmarks, a technical report, model size, licensing terms or a date for wider access, and the demos were selected by the company. Its statements about data efficiency, training hardware and on-device performance could not be checked independently. What the launch does show is that Block, better known for Square and Cash App, is backing speech-native training, a different route from the text-first design behind most current voice assistants.