Roon Radio 1.6 Feedback Thread


(Louis Richardson) #229

Same problem here this morning. Roon radio reverts to library only.

I want my radio back. I don’t like listening to albums anymore :laughing:


#230

I have to admit that I am finding (or not) the same.

I know “the” improvements are in there somewhere based on all of the good feedback, but I haven’t bumped into any yet.

Just an observation, maybe a video with “look here and here” might assist? Sometimes the menus are not shouting the options that are available (aka I don’t click on the dots on every screen).


(Steve) #231

Can anybody here answer the question I asked yesterday in post 193?


#232

Radio selection for my Classical seem ok. My concern is that ROON says it is taking from my library only, but the switch to employ that rule is not on.

Why isn’t it going out to Tidal for Classical selections?!?


(wydi) #233

I think Roon has some technical issues right now


(wydi) #234

I don’t think it’s possible to exlude your own library in Radio mode. Haven’t found a setting to accomplish that


#235

Can anyone at ROON ( or in the know) confirm if this is true? @brian


#237

Right, just seeing this. Thanks.


(Brian Luczkiewicz) #238

Rap and Classical are very different situations.

Classical music has ample of listening volume, but the data that classical listeners generate are not very discriminatory–meaning, most people who listen to classical listen to a broad variety of it, which forms stronger relationships across periods/forms/instrumentation/etc than you are probably expecting.

This is different from virtually all other genres, where listeners tend to sub-specialize. The algorithm also learns that and reflects it back. Because our classical listeners have similar tendencies towards well-rounded-ness, the algorithm is going to learn that behavior and pick more broadly within classical genres.

This is definitely true of my classical listening–I have some favorites from each time period that I keep coming back to and favorite composers that I explore more deeply, but my listening history is going to inform “relatedness” between Chopin’s Waltzes and Nocturnes, Mahler’s Symphonies, Beethoven’s Piano Sonatas, Bartok’s String Quartets, Barber’s Piano Sonata, and Bach’s Fugues.

Rap is a very different situation–it has less than 10% of the listening volume of a more popular genre like Jazz in our member base. This makes the data noisier, especially when you try to start radio from a release that has barely been played yet. It is generally true across the board that content that has been played more makes a better radio seed, because the algorithm has more people from which to learn, but you can feel it much more tangibly starting radio on a very recent release in an unpopular (for our member base) genre.

I do not listen to a ton of Rap, but I have had excellent radio sessions on more well-trodden content that I like like The Low End Theory, Magnificent City, and Eric B & Rakim (which was also one of our case studies during development). I’m sure there is some content at the margins of obscurity, or very new content, which does not make a good radio seed yet.

In part this is about learning to use the tools in front of us well. If you want to hear something like X, and you can think of a few examples, going with one of the more well known options will probably yield better performance, particularly when exploring less popular (among Roon users) genres.

That said, we have about 4x as much data to work with as when this project started in 2017. And we are on track to triple our data set size again by the end of 2019…so this will all get better over time.

One more thought about classical–some of the expectations that you might have are contradictory. Should it stick to the same period? Same instrumentation? Same form? All three? If you’re having such rigid expectations about those kinds of “rails” and judging picks as strictly right/wrong based on whether they follow some imagined guidelines, then you are not really thinking about a radio use case anymore–that is more like a focus + shuffle use case. Radio is supposed to be flexible–it’s not a scalpel.

Towards the end of developing the radio feature, we were thinking about what it would mean to kick off Composer or Composition radio. We had some early prototypes in alpha testing, but no-one took an interest in them or provided any useful feedback about how well they worked. When we realized this, we started to really challenge ourselves on the definitions of what they should do, and while we had good ideas, they had not converged enough by the time we had to release radio for us to feel confident in building them.

One thing we started to realize is that for classical use cases, we didn’t really want radio (as it is currently defined) at all. We wanted something more like “shuffle…but include content outside of my library”. I would happily spend 4-6 hours listening to just Bach or just Chopin, and this is a much more useful click for me on the composer page than trying to play “stuff from composers like this guy”. Sometimes there are clearly related composers, sure, but composers are far more distinct as institutions than artists are in the popular genres–sothe value of “Chopin + Similar Composers” is much smaller than it is for, say, a Jazz or House seed. I could imagine similar logic for other corners of the world, maybe following period/form/instrumentation axes. Lots ot think about. I’m sure some of these ideas will be revisited in the future.


(David Orgel) #239

Thanks for this and other exegeses on Radio. I’m gradually getting a better handle on how this works.


(Anders Vinberg) #240

Thanks, very interesting.
Must have been a fun project.
And really illustrates the challenges of defining “good” when “correct” is meaningless and it’s all about “value”.
Like in my example of the Canon ad from B&H when I was reading the Guardian in Germany. A London or Hamburg dealership would have been meaningless, a Seattle one obvious, but B&H was relevant in spite of being based in New York, where I haven’t lived in fifteen years, because I have shopped online with them.

Anyway, a Roon UI question: I see your point about shuffle being more useful than radio in some cases. But if I also want to go outside my library, how do I that in practice? If I simply pick an artist and Shuffle, I get local stuff, doesn’t include the services. For a performer like The Beatles, Tidal and Qobuz are listed in sections below, and I can Select All for each section (Main, Appearances, Production…) in each service provider and Shuffle over that selection, but I can’t select across several sections, right? Well, I can select several individually, but (a) that’s not scalable, and (b) the sections need to be expanded to a separate page to see all and that invalidates cross-section selection.

And for composers, local, Qobuz and Tidal are separate pages, in fact often multiple pages.

Do I have this right, that there is no way to shuffle across all three sources for content not added to the library?

I never used to care, and I never did Tidal or Spotify playlists, but Radio has changed my view of a quality browse space that is not limited to the mass market. It’s wonderful!

So now I’m interested in more flexible shuffle options across providers, for various selections, like composers and artists (Hélène Grimaud) and labels (EMC). Broader than shuffle, narrower than radio. How would we do that with a non-geeky UI? I’m only asking about the UI, technically it wouldn’t be problem.


(Andrew Cox) #241

I started this thread hoping feedback from users might help Roon develop Radio. It might still do that but it’s also developed into something far more valuable, Brian’s feedback to us about what Radio is doing.

Classical Radio might need a bit more “top down” tuning because of the broader user tolerance. One important aspect of “similarity” for me when I listen to Classical is volume and dynamic content. If I seed Radio from the Moonlight Sonata, I don’t want to hear the 1812 Overture (this hasn’t happened, just choosing an example).


(Anders Vinberg) #242

Ha ha - when I was in college I had listened to Music for Strings, Percussion and Celesta one evening. The next morning, the alarm went off, really tired after a late night but I had an early class and had to get up, I rolled over and turned on the turntable, Bartok starts very quietly and I fell back asleep, and after ten minutes he blew me out of bed.


(David Roberts) #243

First, thanks for the wonderful job on the latest revisions to Roon Radio (and the rest of v1.6 for that matter). It allows me to use Roon much more than prior versions.

Music Discovery with the new Radio feature (plus Qobus/Tidal) is fantastic. It’s got its quirks, but the overall direction seems to meet your goal of making it easy for us to expand our collections. It’s hard to make any final judgments of course, as I expect your datasets to grow and evolve pretty quickly before I really understand the functionality. But I like what I’ve seen so far.

I still sometimes struggle with Roon when it comes to getting a quick “automatic” playlist. And that’s largely because I’m still figuring out how to tell Roon which kind of auto playlist mood I’m in. And also because Roon seems to have 3 distinct approaches to auto playlisting: Roon Radio 1.6 (“Discovery” of new music), Roon Radio 1.5 (“Exploration” of my curated library), and Browse/Focus/ShufflePlay (also “Exploration”). And its slightly more complicated because Roon’s UI for ShufflePlay (and perhaps your data model) limits my ability to operate on tracks separately from albums (I can work on one or the other, but I don’t think I can get, for example, just the tracks I haven’t yet heard this month from just the albums I heard today)

So if I sketch my own various usage scenarios for auto playlists, I can get most of the way there as long as I pick the right core feature to start with, and I can speed things up with tags and bookmarks. But the interaction between tags and bookmarks can still be challenging (tags being mostly static but applying to any library view, bookmarks being dynamic but forcing a particular library view). And it’s all pretty complicated.

But If I could get the interaction among all of these pieces to be a bit tigher with some new controls, I could do it all a lot quicker and I think I’d get better results (for me anyway). I’d like 3 core sets of capabilities:

  1. Guardrails for Radio 1.6. Let me somehow restrict the radio selection on as many dimensions as you’re willing and able (genre, sub-genre, artist, format, resolution, musical style, musical period, instrumentation, form, similarity to some other thing in my library, play frequency or recency, date added or recorded, tag, etc.). For at least some of these (for example, play recency), I’d like to be able to specify the scope to be track, artist, genre, or album. Naming, combining, and saving these guardrails would also help of course. And it would be especially helpful if as you define your guardrails, you get some real-time indicators of what they’re doing to the size of your music universe.
  2. Some way to apply filters/focus by album, then change that to a track view and apply more filters/focus (Mixed Scope selections). This could be anything from: a) track extraction of any (focused) album view into a clean track view that could be re-filtered on other criteria to b) a whole hierarchical set of filtering and browsing tools that allow you to switch among all library views while preserving focus operations on subsets of tracks. There’s probably a convoluted way to do this by first creating a temporary tag and then applying it to a focus, but lordy that’s too much work (and tags being static don’t help).
  3. Let me set a percentage for new music Discovery vs library Exploration in any auto playlist. Say 25% Exploration and 75% Discovery when I’m focused on Roon and want to do lots of track rating. But maybe 10% Discovery and 90% Exploration when I’m having a dinner party and doing other stuff but still want some fresh musical surprises.

My other, likely more frivolous, request would be a way to let the radio seed change (at least if I’m restricting it to my own library) with every song played (“Seed Migration”). Years ago, my Squeezebox setup had an addon called “Spicefly Sugarcube” that used MusicIP profiles to create random walks through a music library, with lots of guardrails. It had an option to use each played song as the seed for the next, which created a delightful way to meander through your library without unduly harsh song transitions. I miss it. :slight_smile:


(Tony) #244

Just a thought. Are there time stamps attached to the data? I certainly have a broad library so I fit that Classical profile but it quickly narrows down in terms of the “mood” of a listening session which may well cut across periods/instrumentation, probably less form. Maybe the discrimination is in terms of the width of listening sessions rather than the widths of libraries?

Just to give a concrete example. It is 2 o’clock in the morning in Denmark and I queued up the Chopin Berceuse. But Radio has queued up two gangbusters symphonies from Mozart and Haydn. I can imagine many having Chopin, Mozart and Haydn in their libraries but maybe hidden in the data is what people tend to listen to after the Berceuse at 2 o’clock in the morning in their timezone?


(Anders Vinberg) #245

Of course I can do that by walking around in each category and each provider, and select all, and add it to a playlist, and after I have built the playlist I can shuffle it.

But that’s work.


(Brian Luczkiewicz) #246

Not sure what the UI will be like, but I want this too, and I’m sure we won’t be the other ones once everyone has digested radio. Sometimes I discover albums via radio, and that’s cool, but when I discover a brand new artist, the very next thing I want to do is an artist shuffle that isn’t limited to library. Composer shuffle also makes a ton of sense.

If we do the “assume Roon were a brand new product, how would we make shuffle work” mental exercise, we would probably just make artist + composer shuffle traverse both in and out of library always.

An option is to do something like what we do to radio and bury a kill switch someplace, like we have for radio:

59%20PM

One of the last remaining frontiers in terms of finishing the UI redesign is re-doing the album/artist/composer/composition screens. There are a lot of problems/evolution issues on these screens at this point, and among other things redesigning them will let us sanitize a lot of sources of potential confusion while making room for new stuff like this.

Part of why it works well to your ear is because our training data comes from our user base, not a mass market user base. The same models will be able to drive other less passive discovery features. People who have more mainstream tastes that don’t overlap with our user base might disagree.

This idea is basically a request to “give us the ability to box the algorithm into a corner where it doesn’t perform well.” It seems great in theory, but the practical realities are rough (we know; we tried stuff like this–it’s much, much harder than it seems at first when the idea first pops into your head).

Lets say you force the proportion to 50% library, but there aren’t enough good options in the library to meet that–what happens? The system either ignores your constraint (confusing…people will complain) or makes bad picks (quality issue…people will complain).

We experimented with a control like this for radio, and it didn’t work very well. Caused both of the above problems. In the beginning, we were going to try and find a way to make the new algorithm work against the library–that attempt failed too. In the end, we have to make a good, defensible, comprehensible experience.

If we add lots of knobs they need to work just as well as the unadulterated version. It seems like a simple ask, but increasing the configuration permutations massively increases the difficulty level of the undertaking. There is also a natural bias against adding knobs to a product design because the more knobs we create, the less people use them. Knobs also greatly complicate the problem of assessing algorithm performance over time to feed improvements back in–because now everyone is running their own tweaked version. Lots of stuff to consider.

We have been playing with an idea sort of like this, but outside of the context of radio.

We call this idea “universal focus” and have spent a lot of time teasing out details. It’s incredibly complex, and also powerful. I’m not sure when it will happen–it’s very cool product stuff, but also extremely costly/time-consuming, and past experience suggests that this is the kind of feature that would never pay for the effort required to build it. So if we build it, it probably won’t be driven by business motivations…it would be because we really felt like we had nailed something innovative and wanted to share it.

Yes there are. Play history (content + timestamp) is already a primary ingredient in this system.


(Gary Proudfoot) #247

Radio has mostly worked very well for me but with 1.6 it’s just going off on completely wild tangents with its choice of tracks.


(Tony) #248

I think that is what I was trying to say. If it were technically feasible to weight the distance between picks in the play history of the entire roon user base, then that would include both library and streaming catalog. There may well be very good reasons why it is simply not the case but I would hope that collective wisdom could be mined for all sorts of interesting radio streams.

I really like the idea of an extended shuffle outside my library as well. But I think radio offers an advantage, or at least something different. To me it is about delegating listening choices to a trusted source, whether that is a DJ, a station, the scheduling of a summer season (like the "Proms), even the house style of a label. I rather suspect that the trusted source of the roon user base would be able to keep any radio stream on track whilst throwing up lots of hidden gems.

Radio 1.6 is doing a great job in so many genres. Maybe a few tweaks will bring other genres into that orbit. Of course I understand if the technical/cost challenges don’t make any sense.


(David Mohr) #249

I now tried radio with a lot of different genres, all seem to have the same problem. The same albums get played all the time, radio is useless for me. After trying Pop/Rock i tried Folk, and as with Pop/Rock I get the same albums played over and over when I start the radio. I’m already tired of Robert Plant, Sting, Valerie June, Rhiannon Giddens and Lucinda Williams. Over and over.