I suspect they are using Google’s machine learning cloud service for at least part of this in conjunction with their own data analytics. Machine learning services still have a long way to go as well, but every iteration of them gets better and it beats trying to construct these services yourself when that is not your core business.
I feel like I am in minority here, but the radio function has never worked better and more accurate for me, go figure… Can’t say I have a very mainstream taste either, but to me the radio function is bringing me eclectic choices which totally fit. Sometimes I don’t like a track or two, but that is to be expected.
1 Like
Interesting. Maybe it would help if the “learning” was enhanced by reports from users of overall success or failure.
No, I suspect you represent the majority. There are around 20,000 users on the Community (of ~100,000 total users). We see about 25-30 people who have an opinion that things are worse and there have been 1900 views of this post. People tend to post more when they are pissed. Not so much when things are good or they have no opinion.
We don’t know if the algorithm was changed or not, but I can imagine that if changes were made in an effort to improve things for most people, some folks at the other end of the curve could see things get worse. That certainly is not good for those people, so it’s important for them to voice an opinion!
Roon reads this stuff so they have the feedback. More importantly, they have the data of how users interact with Roon Radio. Their definition of success as stated by Brian was (roughly), people use it, Roon doesn’t get lots of bad feedback from the interaction with RR, and people play new music they haven’t played before. If their own criteria are being met, they have objective ‘big’ data. If the criteria aren’t met, they know they have to make adjustments.
1 Like
or false expectation
1 Like
I am repeating myself here. For years I have been suggesting a user controlled Radio function that offered a “slider” or continuum based upon how far afield the subscriber wishes Radio to drift from the seed track or artist. The defunct service MOG offered such control, so I know it is possible. It, in my opinion, worked great. See description below:
“MOG Radio… generated a continuous play queue based on the artist chosen by the user. By adjusting a slider within the MOG player between Artist Only and Similar Artists , the user determined whether the radio plays only songs by the selected artist, or whether and how often songs by what the application determined to be “similar artists” were added to the queue. When a user’s song selection ends, MOG Radio begins to play and continued until the user makes another selection.”
What we now have is an algorithm that few understand. We are told it “learns”, ultimately providing the best result. I am yet to be convinced. Why not provide direct user control?
2 Likes
While this might be true to some extent, others have succeeded in developing pretty good playlist generator algorithms. Spotify does an impressive job both at radio and at recommendations. Roon is years behind them.
What I don’t get is that since Roon uses a noSQL database the algorithm doesn’t seem to navigate nodes and edges more than superficially. Artists that appear in the “similar to” section of the UI are usually the only ones chosen for follow up music. Using nodes and edges it should be possible to write an algorithm that selects a random edge from the seed node, plays something from the node it lands on, follow a random edge that doesn’t lead back to the seed node, do this X times, go back to the original node and repeat with a different edge than the first one and avoid the previous X follow up nodes for Y iterations.
Building on that and storing the individual user’s dislikes and reusing them to reject generated choices would probably turn out a much better result than what we get now.
Shouldn’t be too hard to code.
Spotify swallowed echonest which analyses music in order to ascertain musical properties and uses that to inform their playlisting. I imagine since echonest days they’ve added user and community preference to the mix, but the basis of their playlisting is knowing something about the musical attributes of the music, not just a list of similar artists and community listening traits.
1 Like
I’ve been playing a lot of Renaissance and Medieval music lately, and Roon Radio has done quite well in suggesting new albums during listening sessions. Sometimes though it plays tracks completely outside of what I’ve been playing. Totally different mood and genre.
I can’t say if it’s different from before, but I do remember playing Roon Radio for a few days non-stop earlier in the year after selecting Blue Note jazz as the seed Label. I got great music for the whole time before my partner exclaimed ‘I’ve had enough of Jazz now!’ 
“Jazz isn’t dead. It just smells bad.” -Frank Zappa
I used Amazon Music HD ‘stations’ for a couple weeks. I only skipped Miles Davis a couple of times before it stopped playing Miles Davis.
With roon, thumbs down seems to encourage it to play more Miles Davis.
2 Likes
On a related note, I’m not sure that the Discovery algorithm wouldn’t benefit from a bit of tweaking, too…
Led Zep are a Metal band?
A 5 year old covers album is ‘contemporary’?
OK, Rock, but not really hard in my books:
Some party this must have been…
1 Like
There’s an argument to be made about Led Zep. It is unlikely that simple “Rock” or “Classic Rock” would be accepted as a useful genre. The Lou Reed stumps me. I never thought “Hard Rock” was a genre. Not sure about Jeff Wayne - never heard it - can you dance to it?
20 or even 10 years ago, that use of “Contemporary” would be perfectly acceptable. Nowadays? It seems out of touch.
Using the subgenre(s) to label the albums seems more likely to be useful, no?
I mean… there IS space-themed disco out there. More than any of us should care to admit.
Being a metalhead myself, I must say that Led Zep has always been included in the Hard Rock/Heavy Metal canon.
JWMVOTWOTW on the other hand should be classified as symphonic rock. Some classify it as prog, but that’s a step too far for me.
IMHO MusicIP is still unsurpassed, because purely based on the musical properties of a track. It comes up with amazing mixes right across genres, blissfully unaware of metadata or crowd sourced junk.
2 Likes
MOG used audio fingerprinting, powered by Gracenote, to pull genre, mood, style, and other audio characteristics from Gracenote’s database. Roon is philosophically opposed to fingerprinting since it has the potential to slow down the initial import of a user’s library.
instead, they use text matching to match to AMG and other data sources that lack the content-to-content context of Gracenote, TEN, and the old MIP. W/o such context, Roon Radio, and any other discovery mechanism, will always be superficial.
1 Like
Roon aren’t philosophically opposed to audio analysis, they haven’t even dealt with the bug that triggers reanalysis if you change the mod time stamp of an audio file.
There is, however, a (mistaken imho) belief that machine learning and AI leveraging user listening data will scale and outperform the likes of MusicIP. It won’t, because it knows nothing at all about the underlying music.
What are the best players/streaming services out there using Gracenote finger printing?