The story behind the cartoon
Nvidia released Nemotron 3.5 Transcribe on October 28, a free open-weight model that transcribes audio in up to 40 languages. It also translates 31 languages to and from English, and tells up to eight speakers apart in a single recording up to an hour long.
The model runs on a single chip, which is the point of the release: transcription at this scale does not need a data center behind it. A local machine can take a messy recording and return it as separate, readable speaker tracks.
Open weights plus speaker separation is a practical combination for meetings, interviews, and archives, where the hard part was never hearing the words but knowing who said which ones. Forty languages in, eight speakers out, no subscription in between.
The joke is the great unmixer: a hopeless knot of hello in every language goes into one end of a cheerful green box, and out the other come eight tidy transcripts, each in its own pastel lane, stamped FREE. The tangle never stood a chance.

