You give it audio with several people talking, and it tells you who spoke when. The result is a timeline of numbered speakers with a start and end time for each turn. Numbers follow the order people first speak: the first voice it hears becomes speaker 1, the second becomes speaker 2. It does not know anyone's name, so you match the numbers to people yourself. Paired with a speech to text model, that timeline turns a flat transcript into a script with a speaker on every line.
Two things set it apart. It keeps up to eight people apart in one recording, twice as many as NVIDIA's previous open model, and it still works when voices overlap. And it can run live: it only needs to hear between a third of a second and a second of new audio before it labels it, so it can work during a call instead of after it. On a widely used test set that mixes eleven kinds of recordings, it mislabels about a third less audio than that earlier model.
It is small, about 100 million parameters, and was trained mostly on English and Mandarin conversation, with some Hindi and other languages.