Block unveiled Dessa, a voice model trained only on recorded human conversation that talks and listens at the same time, but it's going to safety researchers first.

On Friday, Oct 23, Block, the company behind Square and Cash App, announced Dessa (Block on X, Block).

Block says it was trained "from scratch on nothing but human conversation ... No text, no pretrained models."

Most voice assistants take turns: you talk, they wait, then they answer. Dessa is full-duplex, meaning it listens and speaks at once, and it does backchanneling, the little "mm-hm" and "right" sounds people make while someone else is talking. Block says "Inference runs faster than real time" and that it "runs locally on a MacBook Pro M4 Max or better." Jack Dorsey put it as "small enough to run on your macbook" (Jack Dorsey on X). Block also says it learned from "hundreds of times fewer words" than text-based models and was trained on "only one NVIDIA GB200 NVL72," a single rack of NVIDIA servers. You can't use it yet. Early access goes to safety researchers, and Block put out a call for them (Block on X). There's no release date, and no weights or benchmarks have been published. The Verge listened to demos and found them sometimes awkward or uncanny, including one where Dessa called the speaker "Ronaldo" and another where it went off on a Superman tangent (The Verge).

What to watch: whether Block publishes benchmarks or opens access beyond safety researchers. A voice model that talks naturally and runs on a laptop would matter for local assistants and game characters, but so far it's only demos.