Which package/tool you used for semantic search?
Show HN: I trained an AI model on 120M+ songs from iTunes
211–220 of 444 posts
Re: Show HN: I trained an AI model on 120M+ songs from iTunes
#212Can you do some deduping? A lot of the matches are just the same song appearing on multiple albums or single vs. album version.
Yup sorry about this bug! Need to do some additional post-processing to prevent dupes in the catalog from showing up in the results.
Re: Show HN: I trained an AI model on 120M+ songs from iTunes
#213I find it disheartening that the comments are loaded with people saying how bad it is. The interface is awesome, the ability to have search results pop up and offer suggestions is impressive. I thought the results for a Lana Del Ray song and a song by Carlos Santana were both interesting and suggested new things. We should do a better job of supporting creators here.
Agreed 100%. Folks often forget that a great model is only a tiny part of the equation - strong execution and a solid demo are also critical.
Google/DeepMind might have an RL-based chatbot equal to or even better than ChatGPT, but OpenAI has achieved the most due to its ability to ship.
Re: Show HN: I trained an AI model on 120M+ songs from iTunes
#214I’m not able to paste into the song field, a little frustrating. Otherwise, this has caused me to sign up for Apple Music.
Will 100% improve the UI!
Re: Show HN: I trained an AI model on 120M+ songs from iTunes
#215Is this similar to how Shazam operates? I suppose there's more filtering and denoising going on with mic input.
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.)
Re: Show HN: I trained an AI model on 120M+ songs from iTunes
#216What vector database do you use? Did you run into any scalability challenges there?
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.
Re: Show HN: I trained an AI model on 120M+ songs from iTunes
#217I entered "Spiro - The Vapourer", a brilliant dance music inspired instrumental folk piece with intense, layered melodies.
It recommended "Buzz Cazon - Sentimental Attitude".
I have trouble putting into words how bad that recommendation was. Seriously, just compare those two pieces, do they have ANYTHING in common besides maybe vaguely the tempo?
Re: Show HN: I trained an AI model on 120M+ songs from iTunes
#218I find it disheartening that the comments are loaded with people saying how bad it is. The interface is awesome, the ability to have search results pop up and offer suggestions is impressive. I thought the results for a Lana Del Ray song and a song by Carlos Santana were both interesting and suggested new things. We should do a better job of supporting creators here.
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.
Re: Show HN: I trained an AI model on 120M+ songs from iTunes
#219Fun. I entered one of my own songs and the fifth result was "Love will tear us apart" by performed live by New Order.
Re: Show HN: I trained an AI model on 120M+ songs from iTunes
#220Really cool! From my understanding, it looks like this is doing something like word embeddings and searching for nearby points in the embedding space. Crazy idea: what if you used a dimensionality reduction like t-SNE instead of learning a vector representation? Would you expect similar results?
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).