Block, Jack Dorsey’s company, unveiled a speech model called Dessa on October 23. The unusual bit is in the first line of the announcement: Dessa was trained from scratch on nothing but human conversation. No text, no pretrained models. Sound only.
What comes out is a true speech-to-speech model: audio in, audio out, with no detour through text. It is full-duplex, so it listens and talks at the same time, and it knows when to cut in and when a “yeah” or an “ooh” is enough to show it is still with you. Inference runs faster than real time, locally, on a MacBook Pro M4 Max or better.
What the text detour costs
Block’s case against the usual voice assistants is concrete. Turn speech into text and you throw away everything text cannot carry: stress, pauses, breathing, sarcasm, a voice cracking. Say “I didn’t say she stole my CUDA code” out loud and stress a different word each time—same words, a different meaning every time.
Dessa is supposed to hear all of that natively and produce it too. In the recordings on the project page the model laughs mid-sentence, gasps a quiet “oh my god” at a surprise, and gets louder when asked to speak up. Nobody taught it that, the team writes. One of the recordings is a conversation between Dessa and OpenAI’s GPT-Live-1 voice model.
The training numbers are the real surprise
Language models need staggering amounts of text. Dessa learned language from hundreds of times fewer words, and had to pick up accents, pauses, humour and timing along the way. Speech also never repeats: say a sentence twice and the data comes out different both times, while text is identical no matter who types it. And a single spoken word takes thousands of times more bytes than the same word written down.
Block trained the model on one NVIDIA GB200 NVL72. Behind Dessa sits a small research lab inside Block, founded by former members of Dessa, the Toronto deep learning startup, with six researchers and three interns. It reports directly to Dorsey, and Ragavan Thurairatnam leads it.
The safety argument carries the most weight
Block gives three reasons for the project: curiosity, better understanding between humans and AI, and safety. The third deserves attention. A voice assistant that only sees emotion after transcription has no idea that someone can barely speak through the pain, and keeps going. For a system that takes actions on your behalf, that is not a comfort problem.
Giving a small group of safety researchers early access fits that—and it is needed. A model that reads emotion straight out of sound and writes it straight back has attack surfaces nobody has experience with. Prompt injection by tone of voice is not yet anyone’s threat model.
So far there are recordings and a project page, no weights, no API, no benchmarks. Which means Block picked what you get to hear. The direction is still worth noting, because it cuts against the assumption that the road to good voice AI runs through ever better text models.