While it is amusing to hear Roon radio scrabbling to find show soundtracks in my library it really exposes the weakness in your “genre” system. I am used to find jazz albums as “pop-rock” but to classify Eyal Lovett “Beyond Good and Evil” (a concert recording) as musical theatre is a joke. Maybe you need to recruit users as genre-classifiers. I would do jazz for you at a very reasonable rate
Roon is dependent on the quality of the metadata providers.
I don’t believe Roon software does any categorizations itself.
I have no idea where the information comes from but it is beyond belief that someone would classify the record I quoted as musical theatre. Whatever the cause I have Tchaikovsky and George Formby in the same playlist. It is fun, but is it what we want ? I repeat my offer to take on genre-allocation for jazz (and maybe one or two others) at a very reasonable rate.
The biggest problem with genres in roon is that they are not track based – they are only tied to albums. Getting to track based genres would solve a lot of issues (but not all).
Roon Radio doesn’t use genre to determine what to play; it uses machine learning, i.e. based on your listening preference and those of other Roon users.
Only if you’re streaming. If you’re a pleb using a local library like me you get the earlier Roon Radio experience, which if I’m not mistaken is very much reliant on genre metadata.
When allmusic was still allmusic and owned and curated the database the metadata on the whole was a lot better than its current state under Rovi/TiVo. They’ve become bloody lazy iro quality and completeness.
I don’t know that we know exactly what tech Roon uses for Roon radio. It might be just machine learning, it might not. Having been CEO of a company Gracenote that developed a range of music meta data and technology, we felt strongly that it takes a combination of multiple approaches to be great at radio, playlisting or recommendations. If you use just genres – it’s too broad and you can get bad results. If you just use DSP technology – at Gracenote we had DSP tech that could extract about 30 musical properties from tracks like tempo, song structure, instruments used, etc etc — you could do really great things but it also wasn’t enough. Sonically death metal and christian metal sounds the same but they are very different and shouldn’t be mixed for obvious reasons. If you have both genre and dsp tech you can fix this type of issue because you’d know they are in different genres. BTW at Gracenote we had identified almost 2000 genres of music and editorially tagged these genres to artist. Machine learning brings another great set of connections into the mix but if you just use machine learning, it tends to overweight very popular music that is played by many as related to everything. We were even starting to mine lyrical data in playlisting, radio and recommendations. Some or all of this data and tech was used in media players, consumer electronics and car. Apple as an example licensed all of it to provide a lot of the functionality in Genius.
But I digress… again, we don’t know exactly what Roon is doing but i would hope they are going broader than just machine learning. They should be using genre data also. Machine learning won’t be good enough alone IMHO.
I completely agree with this. I’m friends with the founder of All Music Guide (AMG) who left the company when it was sold to Rovi (now Tivo). They slashed staff and focused just on squeezing more money out of it. The quality, completeness and timeliness of their work is not nearly as good as it used to be. It’s a shame.
Or as I’ve been saying and will continue to say:
Roon is a slave to metadata and not a master of metadata.
Agreed @Jazzfan_NJ . But also getting meta data right is hard. I had about 100 employees working on metadata at Gracenote. Roon could never afford to make that kind of Investement or anywhere near it just for their own use IMHO.
I fully understand the difficulties of getting metadata perfect but Roon also has the user’s metadata, aka file tags, to work with and all too often, as in all the time, Roon throws out good user provided metadata (file tags) in favor of the incorrect metadata provided by their third party sources.
Slave to metadata versus master of metadata
But hey, Roon seems to have resources to program fancy AI for Roon Radio while ignoring the user’s complainants about the mishandling of metadata, which is always answered with “It’s not Roon’s fault, it’s the metadata providers fault.” To which I say “Sorry wrong answer!”
The only viable way to integrate user metadata without the need for a manual data review department at Roonlabs I can think of is channeling it through musicbrainz. But then this already works for basic tag information and if the data at musicbrainz gets better more might get added.
My impression is that a lot of disturbance to the metadata processing comes from the streaming services. From what I’ve seen in qobuz downloads it’s no wonder. As long as the streaming providers prefer quantity over quality this battle for good metadata will not end. They (too?) should employ (more) data curators - like I think Idagio does.
This will improve in time I think as musicbrainz is increasingly leveraged and integrated. It’s a pity that Picard doesn’t offer an easy way to add composer metadata only. That would go a long way to improving exploration without having to rely on Tivo’s lazy ass incomplete metadata. They’re so damn lazy they don’t even get consistency across different releases of the same album.
I apologize for the long post but hope some of you will find this interesting…
Roon has done an amazing job of creating an amazing experience out of most of the music data that is out there but I’m not sure the future will get much better…
I hope I’m wrong but here’s a bit history on this area for context.
Music metadata from labels has been and always will be inconsistent, incomplete and frankly crap. They just don’t care enough to invest in it (other than very small high end labels and/or some classical labels). 20+ years ago when there were these things called record stores, they needed music metadata to integrate into their inventory management, purchasing systems and in store discovery systems. To address this, companies like AMG (All Music Guide) and Muze were formed that addressed this need with large editorial staffs hand entering information off CDs but doing it to a consistent and well thought out set of standards, their own genre structures, and enhancing with artist bios and album reviews. Because there were literally thousands of record stores (from massive chains to smaller stores) there was enough need and enough money to made to support these two company’s efforts to create this data for the most popular music. But they could not cover the entire world of music because there were so many obscure releases from the far flung regions of the world however that it would be impossible to afford it and difficult to even find all the music that is out there.
When the MP3 was invented and the first computer media players came out in the late 90s, you then had the need to have metadata for everything in the world because no one wanted to put a cd in their computer and have it come up with no metadata (e.g. track 1, track 2, track, 3 etc). It was at this time that Gracenote (originally called CDDB) invented the ability for users to submit metadata for cds that had no coverage. Over time, the Gracenote database grew to cover almost everything in the world as it was integrated into iTunes, Winamp, Realplayer etc so the metadata was super broad but not always complete and the highest quality because different users entered data in different ways. Gracenote addressed this as best they could by having editors correct metadata for the most popular music. Even with this process the data from Muze and AMG for the music they covered was higher quality and came with artist bios and reviews that Gracenote didn’t have and therefore was still the metadata of choice for stores, but Gracenote became the metadata of choice for music players. Music Brainz takes almost the same approach as Gracenote with user submitted data but they formed much later to be a competitor to Gracenote.
Switching gears to when digital music stores and then streaming services came to be – they put all the music stores out of business and suddenly most music is being bought/streamed throughout the world from a handful of large companies. Most of these companies no longer wanted to pay Muze or AMG for high quality meta data because it was expensive so most of them have moved to managing their own metadata, taking label feeds and trying to correct and organize them to their own standards (some still license bios and reviews though although many are doing their own now too). If they supported users playing music from their own libraries, they would often still additionally license Gracenote (or Music Brainz) for this so they could get coverage for music they didn’t have in their catalog (while major streaming services have catalogs of millions of releases, Gracenote had data for over 100M releases in their database). This market evolution weakened both AMG and Muze who were then acquired by Rovi (now Tivo) who then consolidated them, cut resources and began producing an inferior product.
That brings us to today. Labels continue to deliver poor data. The Spotifys, iTunes, Qobuzs, Tidals, etc do enough to create metadata acceptable to the masses but can’t afford (or won’t) invest to build the to create the kind of data high end users (like Roon users but we are an extremely niche audience) want – high quality, great coverage, rich data including accurate credits, bios, reviews, photos, lyrics, great classical data, etc. Muze is dead, AMG is diminished and a completely user submitted service like Musiz Brainz and Gracenote can never be complete and high quality to the level high end users and a Roon UI would like because users all submit data to different standards and levels of completeness.
I’m not sure what market forces will move to change this equation. I too would like better music meta data. I understand why users here are frustrated some times with Roon’s metatdata but they cannot afford to solve this equation just for themselves and i’m not sure where the money is for an external company to do it either. I’ll cross my fingers and hope to be wrong about this …
Great post. Having done some work with one of the global industry bodies around rights management and royalties I concur with your view re the labels. Whilst they have a so-called data model and defined standards around metadata for rights management (which if applied could arguably be the gold standard for high quality metadata regarding rights holders) they can’t even get it right among themselves. To this day they exchange information using spreadsheets which aren’t uniform nor adherent to their own standards. Too many egos, too many conflicting interests, too little upside in getting it right - goodness they’d actually have to pay over all that cash they collect but can’t ascribe to rights owners other than the labels themselves. Perhaps the best opportunity to capture good metadata would be in Pro Tools and Abba’s Björn Ulvaeus is at least trying do something to facilitate metadata capture at source.
Leaving the labels behind I’d have to think though that leveraging machine learning and in future AI to enhance music metadata by leveraging the many data sources available on the internet should enable the completeness and consistency of metadata for existing music to be considerably enhanced with no little to no user intervention. Decent metadata is out there in the public domain, but it takes time and effort to sift through it. Applying intelligence to sift and leverage it in enhancing completeness, consistency and accuracy of a resource like musicbrainz should go a long way to creating a very good consolidated source of truth. Distributed computing could be utilised to enable spare CPU cycles of music lovers’ everywhere to be leveraged to accelerate such a project.
Finally and on a completely different note…I’ve been wanting to ask you this for ages… are you the man responsible for killing MusicIP / AmpliFIND by finally turning off musicdns?
The above link is to a discussion on a feature request for adding the ability to display info.txt files similar to way Roon handles/displays image and PDF files. Basically Roon has zero interest in letting Roon users view their own metadata, as in the massive amount of metadata users have in thousands of info files.
Although it may not seem like it from my posts but I do get it that metadata is often not available for many recordings but the metadata is already present in my library in the form of info.txt files, so why can’t these files be easily displayed? (Hint read the post by Brian Luczkiewicz).
Hey @evand. I could tell you countless stories about the lameness of labels – even though there were many great people who I greatly admired. I do agree that if more metadata was captured at the source it would be more complete and you’d hopefully get accurate and complete credits and much better classical data. The link you sent about the new startup “Session” and their focus is very interesting and encouraging. That would be a big step forward. AI and machine learning would certainly help but what we learned at Gracenote is sometimes the most popular and most used metadata for a particular release was actually wrong. You also still have to figure out what genre standard to hook the music too and everyone has their own view of the world and this is akin to religion with many. So it’s very difficult to ever get completely away from some editorial touches and getting consistency across all the players is almost impossible.
And on your last question, no I’m not responsible for killing MusicIP although while at Gracenote I was not sad to see them go…
Agreed re Genres. Being that it is akin to religion there’s never going to be a single answer. The trick is for software like Roon to allow users to apply whatever genre schema they desire, but to apply its own behind the scenes where it may be needed for Roon Radio etc.
Hey - @Jazzfan_NJ
I don’t think it’s fair to say that Roon has zero interest in letting users display their own metadata – they allow you to use your tags versus their data if you choose, but do not read info.txt files which is your issue and is frustrating (and you’re not the only one). I understand that.
I also am a bit a metadata fanatic and hand edited (where needed) my entire 60K tracks of music. I adjusted metadata that was inconsistent (for instance where the name of the artist, album or tracks were all caps versus just the first letter of each word). I also organized them all into a consistent genre structure that I liked at the track level for most of the tracks. The latter was the prime value of what I had done with my collection IMHO.
When I went to Roon I gave all this up. It was not an easy decision. But I found that meta data inconsistencies (like the caps issue above) was less annoying to me in the Roon graphical experience vs the spreadsheet iTunes approach that I was ok with that part. I absolutely missed giving up my track based genres but did so because of the side effects of using them. To use them I’d have to map my genres to roon’s genres (and you can do this) but roon’s structure (AMG’s actually) was much deeper than mine so it wouldn’t map well and I actually like the deep structure. So I gave up track based genres for a deeper structure. I also gave up the work of having to continue to do all the work I was doing on meta data maintenance (or at least much of it) and I like that. I do hope Roon will get to track based genres some time – it would be a big step up.
So I get it’s disappointing that Roon doesn’t support the method you’d like to use to display your metadata but I do believe you may not be satisfied with the results even if you could. Of course this is just my opinion.
Agreed but very tricky…