1.8 Discovery Tools

1.8 arrives with a lot of new discovery tools. The one I like the look of the most is the “Performing the music of . . .” I really like interesting covers of familiar music. Potentially at least, especially with a streaming account, that is a lot of covers playlists with a few clicks. There has already been a mixed reaction to the “In their prime . . .” tool but what do you think about this one? Or any of the others for that matter?

I thought I would try it with some singer/songwriters as that seemed the obvious place to start. I immediately get what on first look (so far) is a great playlist of 50 Joni Mitchell covers. There seems to be something not quite right with the shuffle in that there are several two track sequences of the same artist and there seems to be too many picks from a recent Herbie Hancock / Joni Mitchel covers album. But I have to admit I am enjoying the content.

But the experience so far is not consistent. For example, I also tried with Jim Croce, but roon only pulled up 8 tracks and there was a lot of duplication of the same handful of compositions although covered by different artists. I also tried with more contemporary singer/songwriters like Amos Lee and Jack Johnson but they don’t get a “Performing the music of . . .” banner at all. So I am a bit puzzled by the differences and lack of consistency. Is it the lack of composer credits in the source metadata for example? Or something else. Can the experience be improved?

I hope the topic is of interest. Try and be nice. There are other excellent threads on why my 1.8 is a bloodbath.


The accuracy of the source data seems pretty erratic. I think if you find a performer or composer who interests you, you’ll probably do better to read a good book on the subject, oe even the Wikipedia article, rather than try to understand them via Roon’s auto-constructed pathways.

Erratic… but occasionally magic. Every single ML model I’ve ever been involved in has started this way. Erratic, but just useful enough that it gets some use. By trying it and ignoring it when it’s bad, and using it when it’s interesting, you’re adding to its training set and improving it. It’ll get better faster than you can imagine, just so long as it gets some usage. If it’s so bad that you can’t bear it and neither can anyone else, it won’t get any better ever. So it goes with almost all features in modern distributed applications.


Interesting. What is immediately noticeable though is that with many artists there is no data set at all or a very small one. Why’s that? If you try artists at random with large and small composition catalogues it is very hit and miss that there is even a bootstrap to start with. So there is nothing to train and nothing to learn. My experience with 1.7 was there were very large gaps in the “composer” tags. So is this just not going to work at all with many artists?

Exactly. And even where there are links, they might go away with the next version of the dataset.

That would be my expectation. Johnny seems to think there’s “training” going on, but I kind of doubt it.

That’s a fair question. I’m actually not sure how much any of Roon’s features “phone home”. The aggregate data set in many cases is what is incredibly powerful, but only if it is aggregated. And I’m taking for granted what I would do.

Well, that’s a shame. I’d still like to know how roon is building a composition data set it can draw on to “seed” the ML process in the first place. If that is what is going on.

Roon leverages ML in many areas, but in the OP’s example I doubt there’s any at play, the data set is a straight match of Compositions by X performed by others.

Gosh I sure wish we could crowdsource improved training data then. The amount of person hours that have gone into tagging everyone’s library prior to Roon is kind of insane.

Some artists are not extensively covered by others so they’re not going to have an appreciable list to justify that view as a distinct feature. Of course, this is all reliant on the quality of metadata in Roon’s cloud infrastructure, which I find moved on in leaps and bounds in 1.7 in terms of volume, depth and consistency and continues to improve. Without it many of the discovery features simply would not work. I’ve also found that at a local library level (mine’s extensive) exploration has been greatly enhanced through Roon’s metadata work. The thing that remains missing (and drives me nuts) is the ability to make sense of some of those results by being able to sort them in meaningful ways that facilitate easier exploration - at present you’re mostly presented with a list sorted by popularity or year… that’s ok if the list is 10 entries long, not so if it spans multiples thereof as you’ll no doubt see if you explore Joni Mitchell this way.

Regarding metadata in general, as much as Roon did not set out to become a metadata company, they probably have the most complete, consistent and accurate metadata in the industry at this juncture. Hats off to them on what has no doubt been a hugely complex undertaking to get to this point.

I’m curious how people feel about actively seeking out/discovering new music vs passive discovery whereby it’s presented on a platter. Does the latter make you feel less invested in the artists and/or the music or does it make no difference to you?

Do you still attach meaning to the music you listen to or has it just come to represent an amorphous mass of songs you like to hear depending on mood?

I’m very much of the former, preferring to actively discover new music and it’s interesting to me that whilst the time of the album has passed most “now playing…” posts remain album centric.

I’m afraid that’s Roon’s Achilles’ heel…

I’d argue from personal experience that it is one of their strengths.

You have examples that illustrate your perspective?

Gaps, in the basic Classical metadata coming from Qobuz are not uncommon. Things like Composer, composition identifiers like Opus numbers and with major composers century or more old catalogue numbers. Also hierarchical structuring information for multi-part works that are common in that genre. Better tools to “manually” add value to a streaming account via roon that could improve your overall listening experience was already an issue in 1.7. 1.8 has not improved that in several genres. The existing tools to manually patch the problem have just ended up that little bit more awkward to use.

For streamed content there is no option to use third party tools so it is now just the best you can do with roon’s tools. Traditional groomers transitioning to streaming will want these tools I am sure. Since I got a streaming account about half my library content is Qobuz and as a direct consequence of that I would like better library management tools in roon I never really used very much when I had only a local library.

Roon usage statistics tell me I actually only listen to Classical about a third of the time. I am just as likely to be listening to 80’s indie apparently because that is about a third as well. That sounds about right to me. I would like to be able to build “covers” playlists not for Classical but for these other genres. But the same rules seem to apply as with Classical. If the tracks have no composer metadata or haven’t been identified as a composition it is difficult to see how one cover could be linked to another cover. My initial experiments are that roon’s ability to build a dataset of covers of your favorite artists is a bit hit and miss.

There are obvious hits with artists with long composition catalogues. From my generation Bob Dylan and Joni Mitchel are obvious examples and seem to work. Others less so. Randomly I have tried a newer generation of singer/songwriters like Amos Lee and Jack Johnson and roon does not seem to be able to create a “Playing the music of . . .” banner at all. Joey and Johnny Ramone do better with 7 covers but it is disappointingly thin. If you have a streaming account you can try it yourself with your favorite artists. I don’t know why this is so inconsistent, but I know from 1.7 that composition/composer metadata is often missing or incomplete. This can be a major problem with more contemporary music as the composer credits can be very long indeed so that everyone gets a cut of the royalties. This actually makes matching up contemporary music compositions a much more challenging task than matching up Classical music compositions where there is usually a single composer and maybe an orchestrator or lyricist. I don’t think it works at all with local only libraries. I am not sure.

I would also like to see a “Mixed by . . .” section as that makes a lot more sense with many more contemporary genres. For example, largely because of Qobuz we listen to a lot more really great French music and we stumbled across Møme (Chillwave, Indie Electronic, whatever that means) by accident the other day. That was entirely random from being on the univrsalmusic.fr website for entirely different reasons. It would be great if roon could have got us there instead via a “Mixed by . . .” playlist of another more familiar artist.

I don’t see these sorts of tools as replacing other more traditional sources of discovery. Those traditional discovery tools are not going away especially if you want to drill deeper. But it seems to me that roon, in principle, can open that window on many unfamiliar artists and genres. You can always drill deeper if you wish. With some genres you may have no wish to drill deeper and are content with something more expert listeners may find a bit superficial. I think many of us fall into both those camps simultaneously? I spend time on certain genres but the vast majority much less so. Doesn’t mean I don’t want to enjoy genres and artists I have little knowledge of, even if I don’t want to invest much time either.

1 Like

I can’t comment on the classical side because I’m somewhat of a philistine in this dept in that I search for the performer and then listen to their performances without having much knowledge of who composed and what I’m listening to. With Roon’s recent improvements in this area I can see myself becoming more knowledgeable because it provides a structured approach for me to explore the classical music I do have.

On other genres I find Roon pretty damned good, but I’m local library only. Having said that (and whilst my collection is for the most part very well tagged) I’ve got Roon set to ignore my tagging and use its own, and I see many connections being made between performers and compositions. My gripe there is that Roon presents an unstructured mess rather than a list that one can actually meaningfully explore.

Regarding all the discovery aspects of Roon I like that it makes links between performers and genres etc. within my library, but whenever I turn on Qobuz to see whether I’m missing anything it’s clear NRFY and recommended albums often contains stuff already in my local library and it all quickly begins to feel like this:


Do you have a few artists you would like to hear covers of or think that others might? I could try them for another data point with roon/Qobuz integration. I am curious to see what roon throws up (if anything).

Not off the top of my head, but I can tell you from a quick SQL query of the metadata that underlies my library I have ~236k tracks that appear more than once in my library and many of those will be performances by different artists.

That is pretty much what roon is trying to do with this feature with a much larger dataset, except they are trying to join it to the composer(s) and the track name. Not just the track name.

There has been much talk about Valance and ML and the like, but this seems to me just plain old database stuff. There has to be a data set first, surely before a layer of refinement and ML can be floated on top? My experience with Classical and roon/Qobuz integration is there is a lot of scope for improving the integration at a basic datamodel level in other genres where this may not have been done yet. Once the basics are working I would be curious to see how an ML layer adjusts the results. For example, at the moment, if roon has a deep composition catalogue for an artist it will return 50 covers. Of course it is arbitrary but the question is are those 50 covers rotated, refreshed, static? I don’t know.

I have had no long term experience with discovery in classical. However, I would think that the benefits for classical discovery is not so much in the area of identifying covers but to extend towards similar type of compositions, i.e. exploring into contemporaries or predecessors/succesors starting from something like a Beethoven String Quartet.
Do you have any experience in that direction?

That’s exactly what I did to arrive at that number, match track title and composer