Block's Dessa learned to talk without reading a single word

Dessa skipped text entirely. Block's in-house research lab — led by Ragavan Thurairatnam, who founded and sold a Toronto deep-learning startup of the same name — trained a speech-to-speech model from scratch on recordings of human conversation, with no text corpus and no pretrained model underneath. Block says it picked up language from hundreds of times fewer words than a text LLM, and did it on a single NVIDIA GB200 NVL72 rack.

The payoff is full-duplex: Dessa listens and speaks at once, so it drops a "yeah" while you're still talking, cuts in when it has something, and catches the sarcasm, sighs and laughter that die the moment speech becomes a transcript. Block claims inference faster than real time, running locally on a MacBook Pro M4 Max or better. Thurairatnam calls it "a GPT-3 moment for speech."

The caveats are the usual ones for a company blog post: no paper, no weights, no parameter count, no benchmarks, no price. The demos are Block's own picks, and even those show the model misidentifying a speaker and wandering off topic. It's a research preview — public waitlist, with early access limited to a small group of AI safety researchers applying from an institutional email.

Why it matters: if a text-free model can hold a real conversation, the speech-to-text-to-speech pipeline every voice assistant is built on starts to look like a lossy legacy workaround.