AI/Suggestions or Is the Metadata/Discover aspect getting left behind?

The website for Roon pushes the metadata connections and discoverability as one of the primary features of the app. It’s certainly the primary reason I bought it and I have managed to follow connections between performers to discover some new, great music. With a lot of time and effort on my part.

It doesn’t seem like this aspect is getting much love though. The Discover page is pretty basic and at best elevates content already in my library. It’s removing perhaps a few clicks for me but doesn’t do much more. (It’s also very Sun Ra heavy, for some reason.) It doesn’t look like it considers my listening history, browsing history, ratings, etc. Or if it does it doesn’t look like it leverages those for much of a boost in discovery.

It’s the world of AI and big data now; analysing patterns of behavior and making suggestions. Even if Roon didn’t want to host this functionality themselves they could at least let Tidal pass through their suggestions though I don’t find theirs to be very accurate. I know Roon would do a better job. (You can do a lot with out-of-the-box algorithms these days with high accuracy, fwiw.)

I know you guys do a lot of great work on compatibility with various hardware which would undoubtedly require a lot of QA time and tweaking from devs as 3rd parties make changes but remember, Discovery is what you lead with as your differentiator. Your brand is built on that and your brand is your promise to your users. I think it would be remiss not to continue to expand on that aspect of your service.

Or maybe you already are?


We’re in the middle of a huge project to overhaul this area.


There’s a lot of us waiting with bated breath for these enhancements.

Something to look forward to. May the Gods of bits and bytes speed your way.


I’m cautiously optimistic, but I fear the quest for perfection will get in the way. Roon’s metadata will have to improve considerably for this to bear fruit.

Thanks, @brian. Glad to hear that. Looks like I’m not the only one excited to see what’s in store. As mentioned, I’ve already found a good deal of music I never would have otherwise but I hit a wall through my own browsing. By far my favorite aspect of the app.

I hear you. Obviously I have no insight into how they’re handling this but if they play their AI cards right they could “easily” use the AI to improve the meta data over time if they incorporate machine learning and use their user base’s personal updates as data to train on.

We are working on that too, but I disagree that these are closely related.

When you approach the problem from: starting with all of the music in the world, find:

  • stuff related to x
  • recommendations for me
  • recommendations based on x
  • top stuff in x
  • most loved stuff in x
  • … and other queries like these …

There is a lot of implicit filtering of both bad content and bad metadata inherent to how those queries are going to work. The goal of queries like that is to come up with 10-50 good answers, but there is no objective right or wrong answer so long as the results are useful and serve their purpose.

If something is left out because we don’t know about it, or because poor data discouraged people to listen to it in the past, it doesn’t affect the success of the query much at all, so long as in aggregate, important content is not left out often.

Almost by definition, Roon has good metadata for the post popular (thus important, in the judgement of an algorithm) stuff, so this is not a real problem. ML based discovery almost totally avoids confronting the long tail issues of trying to make metadata for obscure content good and complete.

This is totally unrelated to judgements of Roon’s metadata, which come from an unrealistic comprehensivist viewpoint that could be summarized as:

  • all content should be present in Roon’s database
  • all data should be correct and self-consistent
  • all data should be built to the same degree of completeness
  • all content should be unambiguously and automatically identifiable to a single entry in Roon’s database.

None of those are actually attainable perfectly, no matter what we do or how much effort we burn doing it. We can continue to improve in those areas incrementally, and we will (separate project, separate people, also proceeding in parallel), but it’s not where the low-hanging fruit lies.

I’ve been living with some of our internal demos of the new browsing/discovery stuff for a while. It’s just so totally different than what we have today, and it puts a lot of our current stuff to shame. I can’t wait until we get that stuff into Roon proper and get it released…but still some steps between here and there.