Roon Radio 1.6 Feedback Thread

(Reader of the Internets) #270

I’m thinking it’s all about the genre tags.

The seed album is tagged as “British Folk” and “Folk”. Things like the Smithsonian “Folk Song America”, a 4 CD collection, probably aren’t showing up because it is “unidentified” in Roon, and has no tag “Folk”. Clannad and Siucre are showing up because they’re tagged as Celtic, but also as “British Folk”. Altan, on the other hand, is tagged “Celtic folk”, but not “British Folk”. Joan Baez’ cover of “Let It Be” from “The Best of the Vanguard Years” shows up because that album is tagged “Folk” (somewhat incorrectly, lots of Pop and Top-40 on there). “The Essential Leonard Cohen” is tagged as “Folk”, “Pop/Rock”, and “Singer/Songwriter”, so we got one track from that. The Rosalie Sorrels album also has two tags other than “folk” which aren’t on the seed album. More tags seem to suppress tracks; “The Anthology of American Folk Music”, something like 80 tracks, has more than 20 tags beside “Folk”, so I’ll bet it never is selected. Same for Dylan.

Probably more complicated than I think, but this seems like a good first-level explanation.

(Ged) #271

And the fact that the previous radio was pretty rubbish and is still so and isn’t going to get much better.

(Reader of the Internets) #272

Hey, for me, it’s still the present radio.

(Ged) #273

Well you know how it is then :blush:

(Reader of the Internets) #274

Yes, but… Should it be rubbish?

I’m still trying to figure out what I’ve received from this vendor in exchange for my hard-earned cash. Here’s what Roon said in the 1.4 release announcement (and thus what I was promised when I signed up):

As many of you know, our Radio algorithm has been the subject of much debate since Roon launched. While many users like the selections at first, over time they tended to be predictable or repetitive, repeatedly selecting the same songs, albums, or tracks.

As of today’s release, we have enabled an entirely new Radio algorithm for all of you. Radio 2.0 selections draw on significantly more information, including information about composers, performers, genres, influences, and more, and uses a significantly more sophisticated statistical model for identifying similar content, accounting for content quality, and producing a radio stream that is fresh each time.

Radio will take precautions to ensure the same songs, artists, and albums aren’t picked repeatedly, even if you tend to start Radio with similar content over and over. It will also try to avoid interludes, introductions, podcasts, interviews, and other content which might be intrusive. Note that in Roon, “Radio” means “play more like this” and this is the algorithm that we’ve completely rewritten. “Shuffle” continues to mean “play this group of tracks in a random order” as always.

Supposedly, this 2.0 is what I’m using. Yet I’m seeing continuous repitition of the same artists and albums. The “significantly more sophisticated statistical model for identifying similar content” seems to be working entirely from the buggy genre tags that Roon’s somewhat haphazard metadata has (or does not have, in the case of most Smithsonian collections).

(Ged) #275

I’m not sure why you are still using it either as you seem upset with all aspects.

(Reader of the Internets) #276

Well, never used it much before. Just with all the hoopla about the new radio, I thought I’d try it more thoroughly. To be fair, the seed I was using is pretty obscure (it’s surfaced because it was used as a theme in the Amazon series “Patriot”). I was wondering if they did any actual analysis of the music track, and the apparent answer is “no”.

But this kind of exploration is also a way to judge how well your vendor treats you. Sure, marketing departments always over-promise. But by how much is an interesting thing to know about a company.

(Ged) #277

I was perfectly happy with the previous radio as it did what was reasonable just as I’m happy with the metadata and the like. Roon can only do so much on its own and does things that other software doesn’t.
You either decide that is what you want and go with that, giving improvement suggestions where appropriate or use a patchwork of other solutions. That’s my view on it :blush:

(Reader of the Internets) #278

It is, indeed! I, on the other hand, would like local library radio to work better. My improvement suggestion would be to use actual music analysis like Essentia, which produces stuff like this, which would give the “sophisticated statistical model” something to work with besides the buggy genre tags.

By the way, why must “other solutions” be “a patchwork”? Just curious.

(David Mohr) #279

Tested it again with an Artist radio. My choice was Lana Del Rey. While it is better then the genre radio it still plays a lot of the same albums in every of my 5 tests. Around 50-60% of the songs played in the different tries where the same albums shared in all tests. Again, not bad choices, but i don’t know if this is intented that every time you start a specific radio you get the same albums played with so many music available. Coldplay for example was always part in all test runs, but the only album it played songs from was a live album from Buenos Aires. I don’t like to listen to live music. Coldplay has a bunch of albums out, but all the radio does is playing from that live album all the time.

(Ged) #280

Because they don’t do all the things that roon does in one piece of software.
I can’t see the previous radio getting much attention as the balance point for roon is shifting toward streaming users away from collectors as the type of joining user becomes more mainstream.

(Reader of the Internets) #281

But it’s not clear that “all the things that roon does” are necessary to all users. I would be perfectly happy without MQA unfolding or Roon Display, I think, and probably Radio. Case of featuritis, I’m afraid. What I really want is more work on the core and its weak spots, like metadata.

(Ged) #282

But that’s personal as well for my collection the metadata is perfect.

(Frank Daman) #283

For now I intend to give the Radio algorithm time to learn.

I do however have the same experiences as many of the other posters: very limited variation in artists, much more repetitive than @brian stated in his extensive explanation and once the algorithm latches on to a album, it keeps going back to that same album (like a live album by the Who and one of the “best of” type albums by the Doors).

I get the “safe queue” approach and I agree with it, but for now the algorithm tends to fish in a small pond.

I also wonder if the algorithm uses some kind of hierarchical links between artists (nodes, edges and boundaries). If so, periodically choosing another of the seed artist’s (or genre’s or track’s or whatever’s) edges to branch out from might result in more variety without going off into the deep woods.

If the algorithm only uses metadata, maybe it would be a good idea to include “similar to”-like data in the metadata model (provided this is not already the case).

I specifically chose a Bowie track for my first extended try out, because Bowie should yield a lot of different paths to explore. The results were - as I stated before - a little repetitive given the versatility of the seed track.

On the other hand, the first experiments were done last weekend when everything did a Titanic impression, so I’m not’really convinced that Radio has been given a fair chance to prove itself.

Very curious to see what Radio will become over time…

(Tony) #284

I know it’s no consolation but I had a very similar experience with Radio 1.5. We have a lot of Indie which often seems to end up with a “Folk” tag of one type or another. But we lived in Ireland for 20 years so we also had a lot of Celtic and Dubliners and stuff like that. Radio 1.5 was unlistenable essentially with an Indie seed in that scenario. I don’t seems to be getting that anymore with an Indie seed in 1.6. Has worked very well so far. Also, with Qoboz I am getting very good results with French Pop seeds but that is probably not everyone’s cup of tea.

(Brian Luczkiewicz) #285

First–some framing. Library vs Streaming is the wrong way to think about it–Library has nothing to do with where the files are. You can have a large library from TIDAL and zero files, and you still function in our product like a large library person. Where the content comes from is incidental.

I think the balance is shifting (really…has already shifted, based on our usage data) from files-only collectors to files+streaming collectors. That is the shift that you are feeling reflected in the product with features like Roon Radio.

A steep majority of our users are files+streaming users already. At the same time, we’re seeing the library size statistics slide downwards as we grow as a product and business. We are not quite at the point with the product where someone could walk in with zero library and have a great experience building one up from scratch in Roon, but we will get there. That is an important goalpost in terms of considering ourselves a complete solution.

Lots of people who care about audio are not music collectors first. There is also a batch of relative latecomers to digital music who are introduced to Roon while buying audio hardware. Ten years ago, “rip your CDs” was part of the barrier for entry going digital. Today it’s “log into TIDAL or Qobuz…and maybe rip the few favorites that you can’t find in the streaming services, but don’t bother ripping the other ones.” These people have very different libraries (and disadvantaged experiences in Roon) compared to people like me who have been ripping CDs for 20 years. We are trying to close this gap with some of this work.

The kind of work that you are seeing with radio helps slide the window of “minimum library size required to get a good experience” lower. Other stuff we work on the future will definitely keep this as a goal in mind, in reflection of the actual things that our members are doing with the product.

We are definitely not viewing this as an either or or a shift away from collectors towards streaming users. It’s a reflection of the fact that today’s collectors are streaming users also.

We work on this stuff continuously, and it consumes a serious amount of our R+D budget. It is the only area where we actually have dedicated staff, and it dominates our operations expenses as well.

The problem you’re seeing isn’t “not enough work”, it’s a much more immovable fact of life–improving large, sprawling, mature systems is a lot less efficient than building new stuff. If we spent radio-level effort on metadata over the past year, it would move the needle a lot less for our members. We’ll continue to work on those issues.

One of our most expensive projects last year was completely overhauling Roon’s internal notion of equivalence between artists/albums, which have been the source of the #1 most noisy support topic in the history of metadata and Roon by a large margin. Millions of issues silently repaired in peoples’ libraries, and much less wrong stuff going forward…but not a shiny new feature that gets marketing juice poured over it, so you are not ticking it off in the “effort Roon Labs spent on metadata” column in your mind.

ROON Feature Requests, Problems, & Comments from New User
(Miguel Rode) #286

Couldn’t you switch to the Queue, clear upcoming (if any) and turn on Radio? that would start radio after the song ends and wouldn’t restart the track.

(Miguel Rode) #287

I use (local) radio with Classical a lot. In my experience, the Genre tag weighs heavily in the algorithm. Do you have a good genre classification of your classical works, or do you just use “Classical” for everything? I use a relatively small but effective list of genres and Radio stays reasonably well on the tracks :slight_smile: If you’re interested (and willing to possibly re-tag your collection), I’d be happy to share my list of genres once I get home.

(Tony) #288

Yes, I’d be interested to see the tags you are using. Maybe they would work better. But actually I use a lot of classical sub-genres. But it doesn’t matter much what I seed with, I get a radio stream of predominantly 19th century symphonies and concertos with a few other things thrown in. Do you use a “Classical” tag at all? I use sub-genres but mostly everything will be tagged Classical as well as for example, otherwise you won’t get a composer showing. But I have often thought is that causing a worse problem than it is solving?

So just as an example I would distinguish between solo piano, piano concerto, harpsichord, and organ but I have the feeling this is being rolled up into “keyboard”. I also make a lot of sub-distinctions in vocal music. So I would distinguish between Opera, Oratorio, Choral, Sacred, Secular and Art Song. And even in Art Song I would sub-distinguish German Lieder, French Melodie and English and American Art Song.

I’m not expecting a radio stream ridgidly confined by the tags but if I start with a light vocal piece with so much to choose from I am hoping the mood will be retained with some further light vocal music with the occasional light instrumental piece, maybe a solo violin/cello/guitar/piano transcription. Even mixing it up with a larger orchestral adagio for example makes sense to me as it would be in a similar dynamic range. It is when, no matter what I seed with I am getting the same range of predominantly large scale orchestral pieces that I don’t like it. A similar comment has been made in relation to Pop/Rock and Folk.


@brian I’m afraid this is not the behavior I am getting. Here are some screen shots of my queue when I tried radio a little while ago.

Radio selected two different songs from the same seven albums in the same order before branching out a little bit to a few new artists and then going back to several of the same same artists but different albums and then picking a third track from one of the first albums. I gave up at that point and stopped building the queue.

In case the screen shots are not very clear the original seed was Lumia from the Album Relish (20th Anniversary Edition) by Joan Osborne, one that Roon seems to know quite a bit about.