Block previews Dessa voice model for limited safety testing

Block has previewed Dessa, a speech-to-speech research model that the company says was trained entirely on human speech. Dessa is not publicly available; Block has called for AI safety researchers to test it.

Block says Dessa can run locally on a consumer-grade computer such as a MacBook with an M4 Max chip or better. The company describes the model as full duplex, meaning it can listen and speak at the same time instead of requiring strictly alternating turns.

The company argues that voice systems built around text can discard information carried by speech. It says Dessa can convey cues including laughter, pauses and changes in cadence, and Block personnel have promoted the model as unusually natural and novel.

Those performance and novelty statements remain vendor claims. Block published selected conversation examples, but company-selected demonstrations are not independent tests of conversational quality, reliability or safety. The independently opened reporting provided no benchmark or hands-on evaluation that established how consistently Dessa handles emotion or interruptions.

Block also said it trained Dessa using a single NVIDIA GPU rack. The available reporting did not specify the rack model, model size, training data volume or supported languages.