Block launches Dessa: speech-only, full-duplex speech-to-speech research preview
Block introduced Dessa, claiming a true speech-to-speech full-duplex model trained from scratch on human conversation only—no text pipeline or pretrained base—runnable locally on MacBook Pro M4 Max or better. It is a research preview/waitlist with no public API or weights, prioritizing early access for AI safety researchers.
Block (parent of Square and Cash App) posted Dessa on block.xyz/dessa as “the Conversation Model trained only on speech.” The company says it trained Dessa from scratch on nothing but human conversation—“Sounds! No text, no pretrained models”—without a text intermediate stage or a pretrained base model. It calls Dessa true speech-to-speech (audio in, audio out), full-duplex (listen and talk at once), with backchanneling learned from real conversation; inference faster than real time; and local on a MacBook Pro M4 Max or better. Block has not published parameter counts, public weights, benchmarks, or a public API.
Block frames the project around curiosity, better human–AI communication, and safety: today’s voice stacks often flatten audio to text and drop emotion, sarcasm, laughter, and pauses—“the most powerful AIs in the world can’t hear how you feel.” The page shows live person–Dessa recordings (subtitles added later) and says no one steers what Dessa says or how it sounds in those clips. THE DECODER and TechCrunch restate those claims and note Block’s line that training used a single NVIDIA GB200 NVL72 (vendor/report wording).
Access is “Join the waitlist” under a research preview. A separate call invites a small group of AI safety researchers for early access to hunt unanticipated risks. The lab describes itself as a very small Block research group with roots in the Toronto deep-learning startup Dessa, reporting directly to Jack Dorsey (cofounder, Chairman and Block Head); contact listed as ragavan+dessa@block.xyz, signed by Dessa lead Ragavan Thurairatnam. TechCrunch likewise says there is no public access yet. Superlatives such as “most authentic full-duplex speech model” are recorded as vendor marketing; independent evals and specs remain unavailable.