I may be in the minority here, but having used Qobuz, Spotify, et. al., I find Roon’s daily playlists to be seriously lacking. I can almost never listen to any of them straight through. Take a look at the following image for example. If you’re at all familiar with the group, Pink Martini, you’ll recognize how comical (and jarring) some of these picks would sound bad to back.
Maybe with Harman taking over they can spend some money on improving their radio and playlist creation algorithms.
Wait until you get Bone Gnawers in a Diana Krall playlist. I think Roon is making some progress on this.
They likely can only do so much, because their approach is fundamentally flawed: the dataset they’re working from is, if we’re talking big data, probably relatively small (how many “expert listeners” are creating their own playlists / jumping from one track to the next, and how do you differentiate between a “tasteful” and an “inquiring if I hear the nose-flute on this track as well” style of jump ?).
The one approach that does make sense is Plex’s, which analyzes the tracks in your library, and then does a crossfade to make things flow better. This, for me, has worked infinitely better than Valence in the little time I tried it before giving up, and has been completely devoid of hallucinations like the one you described. Two of the caveats I can think of is that the initial analysis takes A LOT of time / CPU power, and that making it work with streaming services / primarily streaming libraries might be tricky, rights-wise.
I get the size of the user data set being a limiting factor, but wouldn’t you at least incorporate genre information? Given that there are multiple free sources for track metadata–including from sites Roon already displays–it seems they could eliminate the worst train wreck miscues from their mixes.
IIRC, Valence does that (and much more). But it isn’t unlikely there’s either a quality problem with the metadata, hallucinations, or both. All of which Plex’s approach avoids, at least in my experience… it’s also likely that the outliers like Bone Gnawers in the middle of Diana Krall are exceedingly rare… but here again, we don’t know what we don’t know.
The gold standard for me was always Pandora, although I haven’t used them in a few years. And to be fair they licensed the tracks themselves, instead of aggregating them from other providers, which as you say surely complicates things.
I think that having Pink Martini and Tori Amos in the same mix is an order of magnitude different than having Bone Gnawer and Lana Del Rey in a mix.
@Shawn_H_CO This problem is on my radar. The basic explanation here is that this is a mix made for Pink Martini. Pink Martini is a bit eclectic. Eclectic music tends to be hard to generate recommendations around because if you look at what’s around it’s not indicative.
They Might Be Giants for example would be another example where the picks would be challenging.
Also, regarding genres, eclectic music tends to get tagged with fairly indistinct genres. Not that this is true in all cases, but most. ‘Indie Rock’ doesn’t really capture the ride you’re about to go on if you throw on They Might Be Giants.
Just want to say that I really appreciate this thoughtful response overall but the goodheartedness & lightheartedness too. I do wonder what additional data sources might help - is there a way of getting pathways between songs to be smoother or “more variable but in cognitively understandable ways” than going all the way to “full sonic scans” as some competitive products have or MusicGenomeProject as others have? Does really feel like we are in a moment where the expense of using classification models to do zero shot learning is coming down, but I’m no longer in charge of trying to get this kind of stuff done commercially so I hesitate to say “sounds easy” when I don’t frankly know.
The best answer I can give is that I think we have what we need to solve these problems - they are just nontrivial problems.
Even just sonic scans aren’t great when you’re recommending out of the whole wide world of music without the help of in Plex’s case pre-filtering it down to your own music only. The average thing that sounds like something you like sounds a lot like something you like but, well, they aren’t very good.
Yeah not a fan of sonically familiar as it still gives tracks that don’t follow on well, just because it times similar. However their other ways of doing radio do work pretty well but can still repeat.
Great to see Roon have this issue on its Radar and Developers replying to the thread.
FWIW, what has surprised me is the randomness of the quality of Roon radio results. Sometimes, it is delightful! The tracks fit well, they remain in the same Genre, and make me discover new artists and albums. At other times, they are truly disparate.
In any case, I can fully understand how hard to solve this problem is.
A few months ago I received daily mixes that were really good, like e.g. a Cat Power mix where every track was a hit and fit into the mix. I saved that one as a personal playlist and am still putting it on once in a while.
Yesterday I saw a new Cat Power mix where half of the content made no sense, the worst offender being an AC/DC track from Back in Black. I saw it during the day on ARC and wanted to save it for you, but unfortunately it was gone by the time I came home
Not to pull back the curtain too far (especially since what’s behind it is subject to change) but the way that mixes are driven by your play history attempts to make the mix more relevant to you such that a mix isn’t just ‘here are some things similar to Cat Power’.
This falls down if your play history has been dissimilar to the mix base, at which point that lean towards you leans too far and makes it fall over. To an extent your loss on Cat Power will be a gain on mixes more similar to AC/DC.
Anyways - fixing this is in the pipe like all the other issues.
Agreed. Tori Amos, and Sarah McLachlan weren’t really the big offenders in my example above, but Radiohead? Imagine going from a Latin ballroom track from Pink Martini into “Creep”, or, “Paranoid Android”. Kind of stretches the boundaries of eclectic.
I do get difficulties when dealing with bands and artists that cross genres, and are hard to classify. As a lifelong TMBG fan, they really are tough to pigeonhole–although humor and intelligent wordplay are what I would seek to match if I were programming around them.
I really wasn’t trying to offend, just pointing out the current solution seems to lag behind some of your competitors. I’m hopeful it will continue to improve.
It’d be impressive if you manage to offend 
It’s good feedback, I appreciate it and I’ve got it in mind for our future planning.
This explains a lot. While I understand (and admire) the intent, I’m not sure this model will ever work well for people whose listening habits have a lot of variability. Just because I sometimes listen to Chet Baker, and at other times Genesis Owusu, I’m not likely to enjoy an imitation of that genre-hopping on a routine basis. I tend to have vastly different listening habits based on mood, time of day, who’s in the room with me.
I guess if you could envision your variability parameter as currently set on 9, maybe dialing it back to a 3, or 4 would keep some of the intended surprise while reducing the most jarring examples. Easy to say, right?
Maybe I’ll never be the best use case for this strategy. One other question though: does this model ingest its own playlists/radio mixes as listening history to factor in for future track choices? That would explain a lot about why Roon Radio seems to sometimes get worse over time; as maybe its oscillations get magnified in a feedback loop.
I’m not likely to enjoy an imitation of that genre-hopping on a routine basis
Nope - and this was a bug implemented in the original model for the solution. I’ve got a plan to fix it, but the practicalities of approaching a holiday season means that it’s not my first priority at the moment.
Maybe I’ll never be the best use case for this strategy.
I’m a harder use case than you and I’m hoping to fix this for myself too, so have hope.
One other question though: does this model ingest its own playlists/radio mixes as listening history to factor in for future track choices?
Not in the way you are expecting. I know why radio seems to go ‘off the rails’. It’s a different problem, and after I’ve got mixes in a consistent spot radio is (planned to be) my next target.
I think in general having a correlation / or taking it into account the overall history of songs listened to by a users to further improve daily mixes, Roon radio etc is a good idea.
To what extent is this currently done?
I always get opera and I hate opera.
