On Tue, 9 Aug 2005, Doug Friend wrote:
I should have said a time-waster instead of useless. (unless you have
a limited set of files) Can you give a list of different
characteristics that you use for classification ?
I'd have to say I disagree with this, but basically, I think this is because we're looking for different things.
I want to be able to lauch PyTone and have it choose a playlist that I won't have to tweak to any great degree, if at all. For an algorythm to be able to do this, it needs to have intelligent, correct attributes about it's library.
Sure. It's another playmode, like random or eg. album-play, something the user decides to select. Not something forced upon the user :)
Pytone itself would be the 'user' of this information. It reads out the characteristics and match this with other songs. Or more simply it could just use the genre tag(s) (of course automaticaly set by another tool in advance) to find simila songs.
If I have visitors at home, I want to play something peacefully and quiet in the background. My prefered way of doing that is simply by selecting 1 or more songs and let pytone play similar songs. No hardrock, no explicit lyrics... :)
Given that two humans rating & scoring music going into the library will likely score subjective characteristics differently, I really doubt a computer system would be effective at this for things other than tempo and maybe instrumental/voice. My larger concern would be false ratings which would cause problems for the playlist engine.
Rate and mood obviously are subjective. Even though one of the paper uses the mood-characteristic for something different, independent of the listener (and automatically deductable).
But there is more to a song than just tempo. I bet someone more specialized than me can come up with tens of characteristics. If you simply define those and let it run on your music collection that eg. has a proper genre set. I bet the neural network build from that is useful for most other people as well. And all the songs that are miscategorized can be fed into the neural network to make it better...
Of course pytone wouldn't use the neural network, but just use the pre-processed characteristics to find similar songs.
And sure the fact that you like this or that song is complementary to all this information. And important to take into consideration as well when queueing songs.
BTW Sony and philips already have something like this implemented in some of the music players and it appears to work pretty well. So it's not science fiction :) Sadly there's nothing available in Open Source.
Kind regards, -- dag wieers, dag@wieers.com, http://dag.wieers.com/ -- [all I want is a warm bed and a kind word and unlimited power]