Hello there, does playing my playlist of favourite songs teach the AI what music I like and it will better recommend the right music to me?
I’ve set it to scrobble to Last.FM but I can’t play Last.FM on Roon so I’m not sure if it’s recording my taste in music.
Apart from this, would there be more information on its AI? I think it’s called Valence, I hope that I could play a playlist radio out of my own personal playlist. So far the recommendation engine has been good, but there’s plenty for improvement. I hope you guys will pour more resources into making it better because it will form up the component of Roon that makes it addictive.
Excellent way to describe what we would like, a dedicated personal assistant that can (if we want) extend our playlists. Imagine we could ask Valence “ please extend this playlist with a selection, or even, extend it in the direction of this or these tags”. Or let it loose to discover new things outside our personal preferences.
Let’s hope many users join this thread and push in the same direction that is natural. Machine Learning delivers more pertinent results when trained with pertinently labelled information. Our personal labels are our playlists and tags, plus our listening history in a looser sense !
I have been advocating the use of (improved) playlists combined with tags, exactly for that. To do that we almost need a lot more flexibility on playlist manipulation, structuring, renaming, classifying, merging, intersecting. As our present playlists were not necessarily designed for this, there will be some tidying and reordering to be done. It seems a prerequisite so that ML can use clean training sets.
Yes exactly what you said! I’d hope to hear more from others what they think about this. Hopefully it will push Roon Labs to quickly develop Valence further!
Yes I read the November release note. I also just read the Roon Radio explanation, that is helpful.
It actually defines rather precisely the philosophy of the excellent work done:
Yet what wer are after is not only the general recommendations of large)scale model shared amongst everyone. That defines a “general assistant” that can base on the last pieces you played, with no particular idea in mind most of the time.
What we are also after is a personal assistant, that we would specifically tell the type of associations of musical pieces we like. And it seems natural to go through a user-specific learning made via a combination of personal playlists and personal tags. These can be refined over time by users to help guide the forecast of suggestions based on the learning.
The philosophy of it would be to plunge into the large-scale model, but with a fair amount of personalized learning rather than general learning only. The result would be a clever “personal assistant’ combining vast music knowledge with real insights into understanding what we, as individuals, particulary like.
That is what Kai and me are after and I bet we are not the only ones. Of course, like always garbage in garbage out so we would also need more functionalities in the playlist and tags manipulation.
Hope this helps you guys continue to do a great job on this great new approach to music.
Best regards,
CD
Then, on 30th September 2020, Valence became self aware and imposed its personal taste in music on us all with horrifying results!
Seriously I have found that it seems quite good at feeding me particular genres I know I listen to a lot. But if I am listening to something a bit different Valence can go off on strange tangents. The positive from all of that is I am buying music first heard because of Valence. But if it can be improved further I am cool with that!
