Jack Dorsey's Block has built a voice model that learned to talk the way babies do: by listening. No text, no pretrained language model underneath — just recorded human conversation, from scratch. The company unveiled Dessa on Friday through its small in-house research lab, and the demos, reported by TechCrunch, are unsettling in the best way: it laughs mid-sentence and finishes your sentence when you blank on a movie title.
Nearly every voice assistant today is a relay: speech recognition turns your voice into text, a language model writes a reply, and a speech engine reads it back. Block's argument is that the text step strips out the sarcasm, the sigh, the timing. "Interacting with current voice AI can just plain suck," the lab wrote on the Dessa page. "The most powerful AIs in the world can't hear how you feel. We think that's ridiculous, leads to a worse experience, and is potentially dangerous."
Dessa skips the relay. Audio in, audio out, and full-duplex — it listens while it talks, so it can cut in or drop a "yeah!" the way people do. It learned language from "hundreds of times fewer words" than text models need, Block says, and trained on a single Nvidia GB200 NVL72 rack. The finished model runs locally on a MacBook Pro M4 Max or better, faster than real time. Dorsey called it "the most authentic full-duplex speech model ever built" on X; the lab's lead, Ragavan Thurairatnam, went further: "This is a GPT3 moment for speech."
Keep the skepticism handy. Block has published no paper, no benchmarks, no parameter count, no API, and no weights — just about twenty demo recordings it selected itself. The Verge found some of them "a little awkward and uncanny"; in one, Dessa opens by mistaking its conversation partner for someone named Ronaldo. The "first" claim has an asterisk: Kyutai's Moshi did full-duplex speech in 2024, though it started from a text model.
Nvidia shipped a small open voice model the same day — built the old way, with text in the middle. Block is betting the text layer is a dead end, and that a laptop-sized model can hear what the giants can't. The lab is giving safety researchers first access before the waitlist opens, which is the right order for a voice that can also impersonate one.
— via TechCrunch