Show HN: I trained an AI model on 120M+ songs from iTunes

Re:Show HN: I trained an AI model on 120M+ songs from iTunes by jointpdf on Thursday February 02, 2023 @10:23PM (#34636751)

Holy cow does this thing have wildly obscure taste in tunes. I plugged in “Work It” — Marie Davidson, and it returned “Beautiful Weather” — Blemow. This is an 8-minute opus of a techno jam from the album Dutch Cow #13—the 10th and final Holy Cow album released in 2018, a true annus mirabilis from Blemow.

IMO my idea for making something like this really cool is to give the user more explainability (why are these two songs similar? according to which factors?), and then more control over search results (brainstorming here, but stuff like an obscurity slider, importance of beat similarity slider, etc.). You can try to extract explainable factors from your embeddings with something like NMF.

(PS—I like the esoteric results. This is cool, good job.)

[ Parent | on HN ]

Re:Show HN: I trained an AI model on 120M+ songs from iTunes by lelandfe on Thursday February 02, 2023 @10:29PM (#34636794)

Yeah, I'm really appreciating that this doesn't have any idea of what "popular" means, so that its recommendations are just based on similarity, as opposed to Spotify's algorithms, which typically seems to use other user's listening habits to influence the same.

It's not always great, though. I plugged in Debris by Lorn, which is a very somber, but still melody-driven song. All of Maroofy's recommendations are just noise scapes or spooky stock library effects: https://www.youtube.com/watch?v=mBCsdIQi7uk

[ Parent | on HN ]