HACKER Q&A
📣 noashavit

AI for Music Discovery


Are there any AI tools for music discovery by genre/related artists/combination of genres and artist the user likes (that last one would be ideal)?

Spotify's DJ is alright, but it tends to loop through the same artists, at least for me. I'm looking for totally new artist discovery based on preferences / filters

  👤 adrianmsmith Accepted Answer ✓
https://www.last.fm/ has the ability to play "radio" via a Spotify Premium account.

It can either track your existing Spotify account as you listen to stuff and learn what you like, or you can just start using it and skip what you don't like and it learns pretty quickly.

It's not AI though, and hasn't changed much in the last few years, but works well, and a lot of people I know (myself included) consider it the best music recommendation system that we know of currently. Certainly leagues ahead of Spotify's recommendations!

👤 gnabgib
There's https://cosine.club/ which finds music by similarity
👤 diggum
Oo! An opp to share a project I built! Please check out ottomusic.ai or the original demo I built at demo.ottmomusic.ai. We were one of the first AI-powered, prompt-based music recommendation service, and I still think far better than Spotify or Amazon’s recent drops.

(For users of the demo site, you can get a Spotify link to the generated playlist in the console. Shhh.)

👤 martinky24
Why's it gotta be AI based? Try out https://www.music-map.com
👤 drakonka
Not sure that it's "AI-based" in the way you mean, but this is why I've stuck with YT Music for years. It's gotten good at guessing what I'll like, and creating auto-playlists to match different moods which are a good mix of tracks I already know and love and new material. I keep hoping they'll take it further - it is literally the one thing I want Google to use all my personal data (listening history) for.
👤 grobgambit
These all largely bullshit IMO.

The best recommendations are from humans who love music and you also share their taste in music.

Once every few years an amazing recommendation bubbles to the surface on Youtube for me but compared to a music lover's top 100 for a certain year? It isn't even close.

There is no reason you need algorithmic efficiency in this domain. 1000 recommendations might even be objectively less enjoyable than 10 recommendations since it will cause you to not explore any single recommendation all that much. Most of my favorite albums were not 1st listen instant favorites. Actually, quite the opposite.

A great album to me is also going to be out of sample and not in the training data of what I have liked in the past. That is a large part of its greatness.