YouTube Music has a serious flaw, can we fix it?
It keeps feeding you songs you already know over and over again, or ones that sound mostly the same:
"Listen Again", "Forgotten Favorites", "Quick Picks": the experience will be like this for a few hours until you're sick of it: [song you know], [song that sounds... just like the last one?].
Can we fix it? Well probably not as I do not have influence over Google, and I have no idea how a company worth trillions shipped something *this* bad.... buuut we can build our own variety!
First I need a test subject, I chose myself:
We will be downloading a list of tracks to use for this experiment; I chose my music playlist — 676 tracks in my case — for processing.
This will take hours depending on the size...
[...]
Now that we have them:
We will split the tracks into 5 second slices and run them all through a specialized music embedding model, a aMuQ-MuLan (mediocre for the sake of demonstration), that, in simple terms, projects every track into latent vector space.
With this data we can actually see which songs and genres cluster together.
Great! We have all this data, so what can we do with it? :
Firstly, we can now have queues that are actually diverse! Either fresh relative to what's been playing, or a gradual change across different genres if you'd prefer. Both are possible.
Secondly, ask for anything: "Romantic", "Chill techno". Because it works with natural language, you can even hook it up to your AI agent and have it manage things for you.
Let's say you've had a rough week and don't want any slow songs that will ruin your mood for the next few days, you've had enough classical music for today but your playlist is full of it, or you have 3 albums but don't want to be stuck with them.
Thirdly, actual discovery: not more of what you already have, but tracks you don't have yet that land close to your taste. Think of it like auto-expanding your queue or discovering on Spotify.
Fourthly, you can use the data however you want:
Compare it with friends, collect data on thousands of people and use machine learning to develop a whole new recommendation algorithm, so you never have to see [song you know], [song that sounds the same] again.
To demonstrate all of this, I have had the courtesy to create a live visualization of my (quite diverse, I'll have you know) library:
