Oh, right, @CrystalGipsy - I had forgotten that, and you’ve mentioned it before. I’m sorry for forgetting that.
I guess the best way to distill it is, they let their eye off the ball because it is definitely way broken when used with streaming services at this point for some reason…
I just listened to the new CSNY Fillmore 1969 Live, and what does Roon Radio play? Andrea Bocelli, “Time to Say Goodbye” Really? Why is Roon sooooo far behind other services?
Can someone give us a reason why this is happening?
The problem continues in 2024: No idea on how a “similar” music is selected. Especcially there is even a crossover into incompatible genres! ROON: please improve the algorithm!
I’m curious. What was the last last song of the CSN&Y set before you got Andrea Bocelli? This is something anyone can try regardless if you have a Spotify account or not. I am just guessing it was the final track on this new Fillmore Live release. That’s, “Find the Cost of Freedom”
These are the recommendations from Spotify based on that track:
I’d say Spotify did quite well in both cases. One thing I do notice is the very large numbers besides some of the tracks. Are they the number of streams? I wonder how those numbers compare with what roon collects and what the numbers need to be to give reasonable results?
Do you have Qobuz or Tidal connected to Roon, did you train the roon radio algorithm and are you sure detailed genre information are assigned to the track from which roon started the radio stream?
With these three conditions I place, I am pretty happy with the algorithm.
I guess yes, and I in general have the feeling the Spotify algorithm was pretty much depending on popularity with their user base and some kind of connection between the songs in general. Did not use it for quite some years, but after some songs the suggestions were either very closely connected to the starting point, or had a tendency towards mainstream which I found very annoying (No, I do not want Lady Gaga or Taylor Swift even if they appear on samplers or OST with songs I like).
Spotify’s recommendations are pretty decent at times, a little surface level for my taste (I don’t use spotify personally) but generally solid from the perspective of a casual music listener. They have a really expansive user base, which is critical to feed good recommendations (in addition to more metadata / relationship driven approaches)
Yeah those numbers are play counts.
Would absolutely love to take a look at Roon’s internal analytics. The user base is way more discerning for the most part, very curious about some comparisons there.
I’ve switched off radio and rarely use it but I was interested in this new CSN&Y Fillmore East Live release so I queued it up from Qobuz and switched on radio to see if there has been any improvement.
Somewhat worse than I remember I would say. Mostly a shuffle of 7 or 8 of the same CSN&Y albums in their various configurations with a little bit of Joni Mitchel thrown in. No Andrea Bocelli, I just got a Rod Stewart track out of nowhere. I would have preferred the Bocelli. It was also spiced up with a few artists vaguely from the CSN&Y heyday like the Beatles, Stones, ELP, Harry Nielson, even Steely Dan but none of them really fit the CSN&Y live set. CSN&Y were a seminal band which would influence many that followed. You can hear that in the Filmore East live set which combines acoustic and electric proto-grunge Neil Young sets. There must be a godzillion covers and influenced bands that would make a more interesting radio set but none of that is in the roon radio stream, at least for me.
Bizarrely roon also found some Johnny Cash from the time and then detoured into C&W like Steve Earle before pulling back and offering up 7 or 8 CSN&Y tracks from albums it had already played.
I’d like to know why radio disappoints. Maybe roon’s data pool is just too small or narrow in demographics and musical interests.
Recommendations are hard yo. “Having” the user data is one thing, doing something with it is a little harder.
I had to build my own system for this. Nothing else was scratching my itch for music discovery.
The narrowness of Roon’s userbase is probably not hugely detrimental, maybe slightly but there’s benefits too.
In what I do, I tend to weight obscure stuff higher to get rid of ‘casual’ listener nonsense anyways.
What training you referring to exactly? Thumbs up or down have zero affect on outcomes overall just the session your in. Valence has and always will have a mind of its own and is more popularity based than you think. It’s initial concept was to learn from Roons user base as they stated we know best so is was supposed to learn by what we play after certain albums/tracks and what other related stuff we may have in our collections and that feeds into others radio. Other than play music not sure how else it can learn. I used it for years it’s learned not much if you ask me. It plays the same old stuff in the same order more often than not. I stopped finding anything new from it years ago. When it goes off it goes off really badly to. It’s even worse when limited to own tracks.
Oh I think the lack of diversity in its user base tastes is part of the problem it’s so poor, the rest feels like it’s bad design decisions and bugs. It’s been in a bad state for years and seems like a toy that the creator fell out of love with when it got its own mind. Something that seems to happen to Roon features that don’t make the grade often. They are left on the vine to wither and die. Radio had great promise, it started well and then got stuck in a rut playing and repeating same tracks and artists, and thus has never really moved on.
We seem to get a bit of lip service every now and then about some tweak or other but yet to see any fruits of this labour. Much like the promised search improvements that were bandied about years ago almost. Not seen those yet.
Truth be known I think Roon abandoned development of it not that long after release. Either the original code wasn’t up to the job or as you say the data set it had to work with didn’t deliver the necessary results for many listeners. It is switched off for me too now but my wife still uses it until it goes off the rails after about half a dozen tracks.
Any news from the Roon folks about any work being done to improve the Radio function?
One thought - if I use Qobuz or Tidal via Roon, can Roon work a deal with those services to integrate their recommendations software - and only for playing tracks on their services, and not a local library?
Roon should have excellent analytics data on users and their listening habits (if they aren’t capturing it I’d be very surprised).
Having said that, actually doing something useful with that isn’t always straight forward.
I could ramble about this stuff (and how other services have implemented recommendations) all day to be honest.
I think the biggest 2 takeaways I’ve had, from trying to wrap my head around how services do this (there’s some very insightful conference presentations about this) and where they’re getting it wrong are:
A) when you’re ‘directly’ profiting off of the platform’s music material (so Spotify, Apple Music, Qobuz, Tidal, Deezer, etc) there’s a conflict of interest that must be balanced. I think conflict of interest is a loaded term, I’m merely saying that the recommendations and what you choose to show a user (that’s a whole thing in and of itself, absolutely fascinating, look up multi armed bandits if you’re curious.The Netflix guys have some very interesting presentations on that ) must take into account other factors related to the business itself, and the desires of the major label’s providing the content to the DSP. The reason I bring this up is that Roon is in an interesting situation, they’re not the content provider, they aren’t a traditional DSP.
B) ML guys are really smart but they’re often not music savants.
As an example ages ago Spotify was literally using Word2Vec on music metadata for similarity matching apparently. It sounds pretty good, and the surface level results are satisfying, they ‘look correct’ but it’s the wrong approach and it’s an approach that to me, says ‘I don’t really get music’
Spotify on the surface looks like it gets you but like Roon, after a short period it plateaus and then it’s really predictable and repeats again and again. For me Tidal has been the best for suggesting new music and for radio after an album finishes.
Tidal is the only one I have had experience of as I use it when out of the house. It does a very good (and sometimes very predictable) job when an album finishes but it doesn’t seem to fall into the typical Roon trap of causing me to think ‘why on earth did it pick that?’. With Tidal I can generally just carry on listening but with Roon I would have to intervene and select another album to play. Skipping rarely works as it has lost the plot by then.
Agreed, Tidal Radio is spot on and keeps the same mood going, where Roon radio starts meandering off unto completely unrelated music. I used to feel it did a pretty good job, but I am usually disappointed as I have to intervene as well. If only we could just let tidal’s algorithm play through Roon.