Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

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

maroofy.com

Post reply on HN
Print view View original

391–400 of 444 posts

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

#391

Post by scarface74 »

pbronez wrote: post #347
toomanyrichies wrote: post #344

Finding sources for input data is something I struggle with when building deep learning models. Out of curiosity, how did you go about programmatically accessing the music files for all 120M+ songs, in order to create your embedding vector? I can't imagine iTunes has an API which would let a person do that.

Good reminder of the value of Adversarial Interoperability https://www.eff.org/deeplinks/2019/10/adversarial-interopera...

If by”adversarial” you mean a publicly documented and freely available API that has been around in some form for two decades.

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

#392

Post by AlecSchueler »

This is actually great, one of the most promising recommendation algorithms I've come across.

I love that it's working by sound. So often eg Spotify will insist I check out other bands in the same "scene" or from the same era as other artists I like, and in genres as broad as "70s rock" it can be really tiring.

One of the first tracks I tried was Natural Woman by Carole King. I love that it recommended other slow but rhythmic piano vamps with tender vocals by artists singing in other languages, some modern, some old, as well as some sung by men. It even recommended me more Carole King which is I guess shows it's picking up on something constant in her music.

What impressed me was that it recommended quite a lot of numbers in the same key! It was funny clicking through them and the tonal centre being unchanged. It was like they really were different but the same.

It definitely doesn't understand everything about tonality though! I tried some atonal music next, Naama for solo harpsichord by Iannis Xenakis, expecting to get more atonality back. Nope, first result was very firmly tonal: "Suite in E flat major" by Bach for solo harpsichord. It definitely got some essential aspects of the pieces down but completely missed the central concepts underpinning their musicality.

Very promising like I say, can't wait to try more things out!

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

#393

Post by bmlzootown »

subtech wrote: post #254

I just wanted to quickly THANK EVERYONE for taking the time to check this project out and give your feedback! I honestly didn't expect this project to get this much traffic -- I really can't express my emotions via text rn lol. I'm working on an improved AI model that should address a lot of the shortcomings of the current one, along with a lot of other features people have mentioned (playlists, deduplicating results…

Volume control is a must. I had my headphones on, at normal volume, and... let's just say I didn't realize just how loud it would be.

Other than that, however, this seems neat. I've been trying to listen to new music lately, so this'll definitely come in handy.

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

#395

Post by mikeponders »

Good work! I'm trying this out. It seems it tries somehow to match keys tempos and percussion characteristics.

Some suggestions:

- Have a relevant/irrelevant button, so users can tag useful music and suggestions that were irrelevant. Save that data and use it to make your model better, either in realtime or incrementally.

- Allow some other options for sorting tracks also. Have the most relevant but also the most popular (as in charting) on top. Or maybe sort by the user rankings (how relevant they think a track is).

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

#397

Post by bogwog »

I wonder why it doesn't find Weird Al songs. If you search for Michael Jackson's "Bad", I'd expect to see Weird Al's "Fat", but neither one appears in the other's results. Instead, the actual results are interesting, because you can hear some similarities if you listen closely. But they're not what I'd consider "similar"

Still, this is a really cool project and I'm sure there's a lot of potential for building stuff with it.

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

#398

Post by dmitriid »

macrolime wrote: post #355
dmitriid wrote: post #322

Earlier quoted context omitted.

They were acquired by Spotify, and there's been some work done by/for Spotify since then. I'm not at liberty to say what, sadly, as I work for Spotify. I think I can say that one of the main challenges is running this analysis for users. It's prohibitively expensive (or was prohibitively expensive) to use this to keep track of and run recommendations for what users are listening for each user. It can be used on small…

Can you say why Spotify's recommendations are so bad? Something like what OP has made should have been relatively simple to make for Spotify for many, many years already, yet that hasn't happen. Is the whole system just rigged to only recommended a few "sponspored" artists?

Because, as I said above, it's a very complex problem :)

I honestly don't know much about recommendations (and what I know I probably cannot tell). But there's definitely continuous work done on them. But it can also be hampered by extremely conflicting requirements (where "some" both means double-digit procent of users and these "some"s overlap with each other):

- some users want more of the same, some users want a more diverse listening experience. Some of these users are the same user, but on different days

- some users mostly prefer curated suggestions, some users want ranodm stuff. They can also be the same user :)

- some users a heavily weigted to only a few artists, some users listen to evereything and anything. And even this can be the same user :)

- there's probably stuff about licensing, availability, contracts etc. at play as well, because in streaming services it's always there, in very bizarre ways

Basically every single tweak to recommendations will break them. And yeah, Spotify employees will complain about this more than anyone else, all the time :)

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

#400

Post by agentwiggles »

AlecSchueler wrote: post #392

This is actually great, one of the most promising recommendation algorithms I've come across. I love that it's working by sound. So often eg Spotify will insist I check out other bands in the same "scene" or from the same era as other artists I like, and in genres as broad as "70s rock" it can be really tiring. One of the first tracks I tried was Natural Woman by Carole King. I love that it recommended other slow but…

I'm enjoying it a lot too, I've always been a bit frustrated by what you describe with Spotify. It's like it's keying more on genre than sound, which has both pros and cons, but ends up giving me a lot of music that is nominally in the same genre but missing the qualities I like about a certain song.

Probably my favorite song of all time is Close to the Edge, by Yes. Spotify will happily provide me with tons of recommendations for 70s prog - much of which I love too, but some of it leaves me cold.

Maroofy came up with "Good Day" by Leigh Ashford, a song I had never heard of. It's very interesting to compare it to CttE - it's not a very similar song in most respects, but has a similarly prominent bouncy bassline. I like the song, and I don't think I'd have found it via spotify or any of my other usual music discovery sources.

Overall this is very cool, great project.

Post reply on HN