Block has unveiled Dessa, a speech model built on an unusual idea: it learned to talk by listening, and nothing else. The company says Dessa was trained from scratch on nothing but human conversation — no text, no pretrained models — and it is the first model of its kind that can hold a realistic live conversation.
Most voice assistants work in stages: speech becomes text, text goes through a language model, and the answer is read back aloud. Dessa skips the middle steps entirely. It is a true speech-to-speech model — audio in, audio out — and it is full-duplex, meaning it listens while it talks. That lets it cut in, laugh mid-sentence, and backchannel with a “yeah” or “got it” the way people do.
Block says that matters because text strips out information that only sound carries: emotion, sarcasm, laughter, pauses, and a voice cracking with sadness. Dessa, the company says, hears and produces all of it, and its replies arrive faster than real time. Remarkably, the model runs locally on a MacBook Pro with an M4 Max chip or better, and Block says it was trained on a single NVIDIA GB200 NVL72 system, learning language from hundreds of times fewer words than text-based models need.
There are limits to what is public so far. Dessa is a research preview: Block has published demo recordings, opened a public waitlist, and is giving a small group of AI safety researchers early access to hunt for risks. It has not published benchmarks, pricing, or model weights, so the claims rest on Block’s own demos for now. Still, for builders of voice products — from ElevenLabs to Deepgram and Descript — Dessa is a signal that speech-native models are becoming a serious alternative to text-first pipelines.
