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

261–270 of 444 posts

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

#261

Post by raffraffraff »

nxpnsv wrote: post #234
raffraffraff wrote: post #204

I'm wondering if anyone has done something similar, but instead of trying to find similarities in the raw audio, they use tags available from sources like Last.FM, Musicbrainz, Discogs etc? And the ultimate answer to that is probably "those sources kinda suck". Discogs is like a trainspotter on the spectrum, fascinated by release IDs. Musicbrainz is kinda similar (each song will have a dozen matches of wildly differe…

Discogs is great, it just doesn't concern it self with how the music sounds...

Which is unfortunate because it has (on a tiny number of releases) instruments and vocal tags. It's just so unreliable. AllMusic is another decent source for tags, but not instruments. It's the age-old problem with ML/AI: data quality. Garbage in, garbage out. If only we could crowd-sourcev listeners and get them to tag music from a list of available moods, instruments etc. Oh wait .. that's exactly the feature that recommendations services have been removing for the last 10 years.

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

#264

Post by majikandy »

Wow. The songs it suggests have such a similar tempo and vibe that when I try it with songs I don’t know I can’t really remember which is which.

I wonder if this app has more potential for the record companies and songwriters in terms of finding copyright infringement than it does for the consumer finding new music they like?

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

#265

Post by ttctciyf »

subtech wrote: post #236
Ensorceled wrote: post #72

Interesting, I put in a bunch of my favourite songs and found nothing I liked. The songs seem to be over matching on drum beat and tempo so there is a similarity to the song I suggested and the matches but it's often superficial. For instance I picked a song with a very strong snare drum line. All the suggestions also had a strong snare drum line but wildly different melodies, genre's, tone, etc.

Yea, this is due to a shortcoming in the current model's design. Got some stuff in the works for an improved model, hopefully will be able to ship it soon!

It would be great to be able to interrogate the model. For example, I'd love to know what it found to be similar between Sound Chaser[1] and Mellotree Park[2] !

1: https://www.youtube.com/watch?v=Eks6KcV2ufg

2: https://www.youtube.com/watch?v=IYPfbX0DmrE

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

#267

Post by thom »

Nice snappy UI. I found a couple of times on iOS that after I'd clicked through to Apple Music, if I returned, the app had somehow stopped being able to play again but this is very well done. I think I would give a boost to song names over bands/albums in the search as I think a fuzzy song match is probably more likely what a user wants than an exact album/band match on something they've not heard of. I'd definitely use a playlist feature that just queues up the top 20 matches in Apple Music, but I don't know what the API looks like. Anyway, only feeding back because it's great.

I can certainly see what the model is getting at each time, and I've not hated any of its suggestions so far, but I've also not stumbled over any new favourites yet. I don't know what kind of features the model is able to learn, I think it might miss one of the things I like most in music, which is not just dynamics, but something that builds tension over time and then blows up. If there are no longer range features like that I'd certainly experiment with them.

I've had a lot of success with Apple's own suggestions (which are admittedly extremely hit or miss), and I've probably grown my collection 1000% in my 30s and 40s after letting music drift away from me in my 20s. There's nothing better than the feeling of a new suggestion and you click through, and there's no artist profile because they're unknown, and they've got like 300 Twitter followers but you love them like a 15 year old. At least once on here I clicked through and found that I was listening to the only song ever recorded by someone, which seemed quite special.

Post reply on HN