N環球 
Block Unveils Dessa, a Full-Duplex Speech Model Trained on Conversation Alone
Block's in-house AI research lab released Dessa, a full-duplex speech-to-speech model it says was trained from scratch on nothing but real human conversation audio — no text, no pretrained models. Early access is limited to a small group of AI safety researchers.
Why it matters
Most voice AI today works by turning your speech into text first — throwing away tone, emotion, and timing in the process, and answering slowly and stiffly. Dessa, from a small AI research lab inside Block, takes a different path: a true full-duplex speech-to-speech model, audio in and audio out, that listens and talks at the same time like a person does. It can interrupt and backchannel ("yeah!", "got it!"), runs faster than real time, and can run locally on a MacBook Pro M4 Max or better.
The striking part is how it was trained: Block says Dessa was trained from scratch on nothing but real human conversation audio — no text, no pretrained models — and describes it as possibly the world's first AI that learned to hold a realistic live conversation purely by listening. It natively hears and produces emotion, sarcasm, laughter, sighs, and a voice cracking with sadness — exactly the information conventional voice pipelines discard.
Full report
Dessa was built by Block's AI research lab, founded by former members of Dessa, the Toronto deep learning startup, and reporting directly to Block cofounder Jack Dorsey. The project lead is Ragavan Thurairatnam. The official page features live recordings of people talking with Dessa: it laughs the moment something is funny, finishes a sentence when the speaker blanks on a movie title, and speaks louder when asked — with, Block stresses, no human steering.
Block says the model was trained on a single NVIDIA GB200 NVL72, learning language from hundreds of times fewer words than large language models (it concedes the comparison isn't fully like-for-like, since LLMs must also learn things like programming languages). The lab cites three motivations: curiosity, better understanding between humans and AI, and safety — arguing that today's voice AI is nearly deaf to the emotion in our voices, which can be dangerous in settings like healthcare, and that Dessa is a step toward fixing that.
Dessa is currently a research preview: Block is giving a small group of AI safety researchers early access to surface risks the team hasn't anticipated. No pricing, public release date, or open weights have been announced. Note that the company has not disclosed parameter counts, dataset size, or benchmark scores, so the technical claims above are Block's own and await independent verification.
💬 Comments
No comments yet. Be the first!