Block, Jack Dorsey's company, has announced Dessa, a voice AI model it says was trained "from scratch on nothing but human conversation."
Dessa is a full-duplex model, meaning it can listen and talk at the same time, the way people do on a phone call, rather than waiting for a clean handoff like a walkie-talkie. Ragavan Thurairatnam, who leads Dessa, shared one sample chat. The exchange has a mostly natural back and forth, with one glitch: the model at first mistakes the speaker for someone named "Ronaldo." Block shared other conversations too, which can be occasionally awkward and uncanny.
The announcement called Dessa the first AI model you can have a realistic conversation with that was trained entirely on speech. That is Block's own claim, and a hard one to verify from a few samples.
The training detail is the interesting part. Many voice assistants work as a relay: transcribe speech to text, let a text-based model write a reply, then read it aloud. A model built only on speech points to a different route, where recorded conversation, not writing, is the raw material. Full-duplex behavior fits that approach, since real conversation is full of overlap, pauses, and talking over each other.
It also means Dessa will presumably inherit the way people actually talk, stumbles included. Calling a stranger by the wrong name is a very human mistake.
For anyone who has felt like they were issuing commands to a voice assistant rather than chatting, that is the bar to watch: whether a conversation feels like a call. Block's own samples sit closer to that than to a command line, with the occasional wrong name.