Voice and diffusion models, trained where the data lives

Resemble AI builds production voice models — systems where "better" is not a vibe but a measurement, and where the training data cannot leave the building.

Their models train against environments built inside arkenOS: evaluators define what a good generation is, runs execute on Resemble's own hardware, and every experiment keeps its full lineage. The loop composes their existing pipeline rather than replacing it.

A detailed write-up, with results, is in progress with the Resemble team.