Machine learning is essentially reliant on meta-data efficacy. Most evident in Roon as the Tags applied (though I’m presuming it also uses integrated Genres and the like), It doesn’t matter how efficient their algorithm is at correlating your preferences from the tag pool if those tags are inaccurate and I’ve seen some really crazy classifications. For example I like a lot of ‘Epic’ content (Two Steps from Hell, Ivan Torrent etc.) and an awful lot get classified as ‘Classical’ because of the primary component instruments so right after a rousing blast of hero-saves-the-day-by-wiping-out-an-enemy-fleet-with-their-pinky I get Clair De Lune (which I love but talk about a cold shower ). Still I can see why and use the correction options. Up a few tracks later though I get Taylor Swift because Jo Blankenbgurg’s “Petrichor” is classified as Pop+Pop/Rock… .
I applaud their work in trying to make the system more intelligent but I think they have mis-prioritised. The meta-data needs to be flawless or it’s just garbage-in/garbage-out.
One improvement I think they can make quickly would be to apply ML as a crowd-source trust system for user-modifying tags. E.g. An option for everyone to reject the current tag(s) and apply new standard ones (not just their own), the ML algorithm evaluates a number of criteria before deciding to update the central classification for the track.
How many users have performed the same action and applied the same tag(s)
Of those users how many have libraries with a high proportion of the recommended tag (adding an experience factor).
Of those users previous updates how many were accepted historically vs. reject.
This is the kind of analysis that ML is built for and providing the algorithm is ok since it’s working on explicit entry history it should become quite accurate over time. The front-end work for Roon is in building this input and analysis method but the long-term work is from the community so it’s cost over time would be minimal.
As a user, I’m not interested in putting in the work to reclassify music. As per the marketing hoopla, I am paying Roon to understand my music. By extension I take this to mean Roon has to understand music.
For me audio analysis based on sonic similarity seems to be the only way forward. It’s up to Roon to make good on their promises.
I didn’t mean to imply this should be shifted to the users as the norm, I think one of the reasons we all buy into Roon is supposed simplicity. I just like to have options under the hood when I need them (instead of heading back to JRiver to do any maintenance).j
While I find all of these posts interesting and sometimes highly informative- one aspect of this topic is conspicuously absent:
Any word at all from ROON. 83 posts to date and zero from ROON.
That is very disappointing. On this (and other subjects) they remain mute.
Fair question! Many have requested Radio to have functionality different and beyond what it is currently. I include myself in that group.
To be clear, I don’t view Radio as either “rubbish” or “great”. It’s somewhere in between. Of course, that means it could be better.
A response to enhancement requests contained here would be beginning of a conversation. Silence is not.
Has many Disco elements and I completely agree with that as one genre classification, symphonic rock is another, and I’d include spoken word.
It certainly was considered disco, and dismissed as such, by everyone in my high school that I played it for. Perhaps, time has a way of distorting genres. Music some people regarded as “so harsh and heavy” when released; now comes across as tame.
I’m not hearing the disco myself and over here in Belgium I don’t think anyone ever categorized it as such. Just goes to show that genre classification is hard to do and may even shift with time.
And I agree with you that what once considered shocking can sound quite tame now. Take Nina Hagen’s Unbeschreiblich Weiblich for instance.
If I look at the genre tags for Thriller, I don’t see disco there (post-disco, yes, whatever that may be). I was a pirate radio deejay at back when the album came out and I can assure you it and all the singles taken out of the album were stored in the disco bins without anyone ever even considering any other genre.
Which leads me in a very roundabout way back to the premise that genre tags are not a good basis for a “radio” generating album.
First about responses. I would prefer Roonies spend their time working on 1.8, not assuaging our incessant curiosity.
About Tastes: I suspect a histogram of a population’s taste (however that gets defined) would have “fat tails”, i.e, very dispersed. For those with centric tastes, the Radio should work well. But the true test of AI and Roon Radio is how well it handle the “tails”. And that aspect of the Radio isn’t quite up to snuff.
That’s a pretty funny statement to make since Roon Labs says Roon Radio picks the songs based on your queue.
Bill_Janssen
(Wigwam wool socks now on asymmetrical isolation feet!)
96
What they mean by that is that they pick the songs based on the tags on the album that the last song in the queue came from. Only the last track counts.
Yes, thanks…I’m aware of that. I am simply not impressed by that. They should analyze the entire queue, not just the last track. That’s what I’ve said in the other thread. Just using the last track is not necessarily representative of what you might have been listening to. My point is that Roon sells Roon Radio as this amazing AI unique wonderful thing…and, to me, it’s not. So, in accordance with this particular thread, I’m simply agreeing…I’m not impressed with Roon Radio. It doesn’t do anything that any other music service offers in that regard. Pandora can create a more appropriate station from what I’ve been listening to than Roon Radio.
Look, I have no issue here other than I don’t think Roon Labs should be touting Roon Radio as this amazing tool when, in my own specific experience, it’s not. Strictly my opinion. Others think differently. That’s fine. Free world and all that (well, at least for now anyway).
That being said, it’s a convenient tool for listening, and I use it. I just hope it gets more intelligent and useful over time.
I’m afraid I have to agree with FastHound. I use it regularly but that doesn’t mean I think it’s very good. I find two cases. It gets it completely wrong on the first track in which case I often go to live radio. It gets it right but starts repeating after 15 to 20 tracks when I will often also go to live radio. Seems at the end of the day I still prefer a lower rez radio stream DJ’d by a human being rather than a higher rez radio stream DJ’d by the roon AI.
I am curious though. Pandora has been mentioned. Are there other AI based radio streams that do better. Gracenote acoustic finger printing has also been mentioned.
Generating a good radio flow isn’t rocket science but it requires knowledge of music.
Back in the pirate radio days we used algorithms for new deejays: stuff like one chart hit followed by one older track with the same tempo followed by a new mid tempo release followed by… and so on. This worked great until they got the hang of it.
My conclusion is that Roon radio will never amount to anything until some formof music analysis is implemented, preferably alongside a few tempo/mood fixed algorithms the user can choose from.
This is what MusicIP did, and it still does a great job many years later. The reason is simple, it knows things about the music in your library and its based on the music itself, not some inconsistently applied, largely incomplete genre metadata.
I think those of us who never worked as DJ’s are instinctively doing something like this when building a playlist or queue. But as you say that does require some “expertise” of the music.
I don’t really know but there have been a few comments from roon about their radio and discovery algorithms. If it is going beyond something simplistic with genres it sounds like at least some of what is going on in the background is an attempt to leverage the “expertise” about music from the listening patterns of the roon user base. Most of us already instinctively do that by regularly visiting the very popular “what are you listening to now threads” to learn from more expert listeners.
But if roon really is trying to automate that instinct somehow from the listening habits they collect I can’t really see that working either. It’s just going to end up as a lowest common denominator it seems to me. I like the sound of the Gracenote acoustic fingerprinting approach and I am sure it’s used in a lot of services but does anyone know of a player or library manager that uses it so that I could experiment on a local library?
). Still I can see why and use the correction options. Up a few tracks later though I get Taylor Swift because Jo Blankenbgurg’s “Petrichor” is classified as Pop+Pop/Rock…
.