AI and music discovery!

The potential for incredible music recommendations is the only thing that doesn't scare the shit out of me about AI. Everything else is...well I'll leave that for Marv's thread let's not make this one political... but lots of power in the wrong hands let's just say.

But what if Apple or Amazon or Google (the only ones who could afford it likely) implemented really good AI to analyze all of their customer''s music libraries and come up with a master formula or just made lots of connections between artists or something and then it could look at your music library and give you incredible music recommendations that you actually like.

I'm trying to find more of a particular kind of metal right now, and put a band into Apple Music and then Amazon music and none of the music played in the "radio station" are remotely to my liking, nor do they really sound like the album I started the station with. Which is what always happens but I keep trying. For me, music recommendations on all streaming services really sucks. I almost never discover music through any of them including Spotify. Pandora should be better, but it only analyzes songs, and every time I use it I find individual songs that I like, but they're often singles that deviate from the band's main sound or just random one offs' on a shit album. I never find bands or albums I like from Pandora.

But AI could really change that and create links between similar artists that we could listen through and they'd actually be similar, not just I don't know - how do they do it now, sometimes I think it's based on cover art or something lol

 

I wonder what algorithm or system Roon uses for Roon Radio? It recommends and plays some killer stuff after the album or song you choose ends. It's the main way I find new bands lately.

 

https://valence.roonlabs.com/info

Edit: I would be interested to learn what algorithms they're actually using for each of the different pieces, but I'm sure they view the "knob twiddling" within each model as proprietary to Roon. I'm also thankful that my day job exposes me to a lot of data science mumbo jumbo so I get to learn the language, as it were. So the technobabble has begun to interest me; Roon's Valence included.

 

chatgpt is not too bad for making basic recommendations and I use it to find similar artist/albums. I'm going through a room trial at the moment and trying to get an idea of how well its algorithms are working.

 

Preface: I don't know the innards of (most) proprietary music recommendation systems but I've worked in the field and I follow it closely. Everything I say here is based on general, published technical knowledge, summarized for your pleasure (or disgust).

The summary: Typical recommendation systems (for any domain, not just music) use two main sources of information: (1) user activity, and (2) metadata. The first recommendation systems where based on collaborative filtering of user activity: basically (snake eating its own tail), it assumes that similar users engage with similar items, allowing user and item similarity functions to be estimated from observed user-item interactions. Over time, system developers realized that most items, especially new ones, don't have much activity, so they don't get recommended, so they don't get enough activity to be recommended (rich-get-richer effect). They started throwing in metadata to fill in the gaps: basically, if Fernando and his similar users seem to seem to listen to many Dave Holland items, they may also like this new Holland item that has no listens yet. Soon enough, artificial neural networks, reinforcement learning, and a lot more got into the act.

However, these methods use very limited information. Collaborative filtering is totally dependent on user interaction, and metadata is restricted to whatever information system developers chose to include in the databases describing items (such as what you might find on Discogs). That seems a lot, but it is shoehorned into a database schema that tends to miss many of the more cultural, social aspects of music: X played with Y in band Z, but they never recorded together; A replaced B in bass guitar on band C; fans hated how A destroyed the trademark rhythm of band C; X's cover of S is surprisingly sad; and so on, an open-ended set of relationships between artists, songs, fans, venues, producers, record labels, ... That's where LLM (large language model) chatbots have an advantage: they aren't restricted to a predefined database schema, and can (try to) connect many kinds of information that may be relevant to music preferences. We'll see whether this can be made to work better than current recommendation systems.