Hacker News Reader

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

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!

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

Sorry! The service is sometimes bugging out due to the recent surge in traffic, which is causing some songs to consistently fail.

Things should become much more reliable and better soon!

Also, sometimes there's a delay between when a song is available in the text search bar, and when it has been indexed with my AI model (which can also cause this error).

This will also be fixed!

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

Thanks! Definitely welcome all the feedback here.

Gonna ship a better, improved model soon!

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

sorry mate, 100% didn't expect this much HN traffic.

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

got it

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

Thanks! :D

I personally think I need to improve the UI a lot, but maybe that's because I've spent way too many hours staring at it lmao.

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

haha that would be fun

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

100% agree with data cleaning. Got ship some updates for this!

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

So many people have asked me the same thing lol. Shipping soon!

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

lool

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

Thanks! Gonna add user accounts with play/like history!!

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

Thanks! Yup, this is basically custom music embeddings + nearest-neighbor vector search.

I personally found that vector representations performed significantly better than other approaches.

And the results will actually be a lot better once I ship a better model (the current one can definitely be improved upon).

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

100% ran into a ton of scalability challenges lol. Maybe I should write a blog post about it sometime.

But for now, ended up using plain old FAISS.

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

Hmm, not quite. Shazam's approach is more optimized for exact matches. Mine tries to be more optimized for similar matches (if that makes sense lol).

So my approach is intentionally a bit fuzzier than something like Shazam, since apart from exact matches, you also ideally want a continuous representation for music similarity (so you can find songs that are 80-90% similar, for example).

That is hard to do for approaches optimized at exact matches (which usually use something like audio fingerprinting, etc.)

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

Will 100% improve the UI!

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

Yup sorry about this bug! Need to do some additional post-processing to prevent dupes in the catalog from showing up in the results.

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

Thanks a lot for your feedback! You're right, I should probably add a better song player UI element.

Also, as I improve the model & add user accounts, recommendations should hopefully get much better and more personalized!

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

Thanks for your feedback!! Yup, model can definitely be improved (and perhaps I should also add a language filter)! Noted! :)

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

Definitely understand the current model's shortcomings that people have mentioned.

I actually think this is largely in part due to how the current model + training process is designed.

I have some ideas for improving things on this front, will give it a try and push an update soon! :)

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

Personally, from using it a lot during development, I found that I kinda developed a sense for which types of songs it'll do really well on, and which types of songs it will sometimes struggle with.

But a lot of this should go away with a better, improved model!

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

Haha, thanks! :)

Will definitely work to improve the current model!

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

Thanks a lot, definitely have a lot of work to do with improving the model!

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

100% agree with a language filter, working on one (along with a better, improved model!)

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

Oof, honestly didn't expect the surge in traffic lol. It should work, as auto-scaling kicks in, etc. Sorry about that!

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

haha didn't expect many people to guess this!

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

Hey thanks for the feedback! I definitely have a lot of improvement to do on the model, it currently performs better for some styles/genres of music than others.

But the model architecture I'm using is kinda outdated as well, gotta iterate on it more to improve it further!

I'm also thinking of letting users upvote/downvote results, which can also help improve quality on the ranking side.