Live data from Hacker News

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

maroofy.com

421–430 of 444 posts

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

#421

Earlier quoted context omitted.

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…

User tagging isn't a panacea either, because people tag inconsistently, and people who tag a lot are probably not very representative. For an extreme example of that, see the boorus. Some machine learning people have become interested in those, since they are huge dataset of extensively tagged material ... or maybe it's the booru people who have become more interested in machine learning. Either way, I'm sure they're…

This is very true. I'd pay for a metadata-only / playlist service that works with Spotify/Tidal/Apple/local music.

And don't allow free-text tags. Instead you give a list of available tags - the lowest number needed to describe most tracks. I mean, let people add their own if they want, but you should ignore those while training the model.

I actually think that instead of trying to tag some specific mood (eg "happy") some sounds be a sliding scale between two opposites:

happysad

Instrument tags are easier to understand. Give a list of instruments (or instrument types, because the user might not know precisely which woodwind or percussion instrument it is) with checkboxes beside each.

Some users will be experts because they play woodwind. Let those users apply to become experts, pass a test, "identify the instrument", and if they pass, give them half price subscription as long as they moderate X tunes per week.

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

#422
post #218

Earlier quoted context omitted.

a better job of supporting creators here You can do that by commenting on the things you liked about the project! But without the meta because that just begets more meta and ends up detracting from the thing being showhn.

I did the latter and think the former is necessary. It is like when Dropbox was on here and everyone shit on it. You can offer criticism without being a jerk.

the former is necessary.

It's not because it does the opposite of what you want. Write good comments, downvote or flag the bad ones, email the mods about egregious stuff. Inveighing against the snapshot state of the thread almost always ends up wrong (it's pretty much already wrong about this thread) and generates noise of its own which works directly against your intended goal - people showcasing their work and getting feedback about it.

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

#423
post #354

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.

Also would like to know. I can't even listen to the full songs, and assuming I have to pay. I can't imagine buying 120 million songs, so it has to be some collab with iTunes.

> so it has to be some collab with iTunes.

There’s no way today’s Apple would allow such a collaboration. They’d just keep the feature and market it as part of Apple Music.

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

#426
post #132

Seems like the recommendations are somewhat hit or miss so far. I would suggest finding a way to say why it recommended a particular song. I know that this is already common, but maybe you could come up with a new spin on it. Also, other comments have suggested that this was trained on 120MM of the audio previews instead of the full songs. That might explain why the recommendations seem a little off for some people.

I'm totally going out on a limb and guessing here - I'm more on the UI/UX side of things so I know nothing low-level about what goes into a building recommendation engine, algorithm, or whatever. But I do know music pretty well, and this feels like it's matching too closely to the technical aspects of the music and not the overall theme/je-ne-sais-quoi that makes a song something you feel . I tried two of my all-time…

Honestly, thanks for taking the time to give such detailed feedback!

The current model I have isn't as good as I want it to be, and I'm working on a newer one with a different training process as well. This should address some of the shortcomings people have mentioned.

I was really nervous of shipping my v1 with the current model, but thought I might as well share what I have so far with the world, in case someone finds it useful lol.

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

#427
post #337
post #325

It seems to be finding snippets of songs similar to snippets. I tried on Metallica - Unforgiven III, which starts off with a slow piano composition, and then enters a riff and cuts out. It ends up recommending piano songs, many Korean ones. There's some interesting ones like Ghost - Cirice, where it finds other songs with similar riffs. I like Ghost's music in general, just not the Satanic themes, so this is a great…

good point, probably a final version should have a pipeline like 1. cluster song segments into styles and 2. search for each cluster or only the main cluster. What would be a good NN architecture for the first step?

It seems like a data problem, not architecture. Usually these preview snippets capture the essence of the song. It works on say, Unforgiven II but not Unforgiven III.

We'll likely see a lot of first generation products tackle these low hanging fruit like "I dumped data into AI/ML and here's the output". But the next generation will likely be people who can handle data better - interpret it, clean out bad samples, and so on.

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

#428
I think you may have inadvertently just closed a loophole with iTunes and will be the one and only non-Apple employee to train an AI model on 120Million songs in the iTunes library. Further, Spotify employees/management will notice the success of this work and ensure it doesn't happen there. Please consider open sourcing your datasets.

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

#429
This is cool. Have you tried Plex? They offer sonic similarity matching that's trained on a user's personal music library. I think it might be fun to do a comparison? This feature was added in 2021 though so I'm sure models have improved since then.

https://www.plex.tv/blog/super-sonic-get-closer-to-your-musi...

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

#430
post #422

Earlier quoted context omitted.

I did the latter and think the former is necessary. It is like when Dropbox was on here and everyone shit on it. You can offer criticism without being a jerk.

the former is necessary. It's not because it does the opposite of what you want. Write good comments, downvote or flag the bad ones, email the mods about egregious stuff. Inveighing against the snapshot state of the thread almost always ends up wrong (it's pretty much already wrong about this thread) and generates noise of its own which works directly against your intended goal - people showcasing their work and gett…

Thanks for sharing your opinion
Post reply on HN