Comment by mac_

A database populated with audio metadata (including a link back to YouTube or Spotify or whatever) that includes vector embeddings for the audio. That way I can grab clips of music I like from YouTube, generate vectors for them, then find similar things in the database.

It's off to a rocky start though, as I've initially populated it with YouTube-8M and AudioSet, neither of which are music-specific. The search results can be... Weird.