[piecerough]

It's very related to LLMs. Though instead of text tokens you are working with audio tokens (e.g. from SoundStream). Then you go to audio corpus, instead of text corpus.