Block released a research preview of Dessa, a speech model from its AI lab that can listen and talk at the same time and runs on-device on a MacBook Pro with an M4 Max chip or better. Block trained Dessa from scratch on recordings of people talking, with no text or pretrained models involved, using a single Nvidia GB200 NVL72 rack, so the model picks up sarcasm, laughter, sighs, and a voice breaking with sadness that voice systems built around text transcripts lose. Dessa isn’t available to the public yet, and Block is letting a handful of AI safety researchers test it first to find risks it hasn’t anticipated. Square already offers AI voice ordering to its sellers, but Block hasn’t said whether Dessa will power it.
