Also, looking for the interest "bhajis loops" on LiveJournal brings up this journal, which may belong to the author, or someone else who is a French researcher/student interested in Bhajis Loops, all things Indian, DSP algorithms and PalmOS programming (not to mention Stereolab and, umm, Architecture In Helsinki). The journal says some interesting things there, from discussions of audio algorithms to observations on why signal-based classification of music is doomed to failure:
Music similarity is to me a cultural feature. For example, the distinction between two genres can be very artificial, and is sometimes based on features like "the way the singer is dressed" or "the amount of sexual contents in the lyrics". An example I like mentioning is the East-coast and West-coast rap. How a computer could tell the difference ? Covers are another example of cultural music similarities. You can totally change the tempo, instrumentation, even the style of a song. But it will still be similar to the original, and will be recognized as a cover of the original song, not as a cover of the original work. A final example is when an artist is inspired by another. For example, a lot of Stereolab fans are comparing their music to some of Steve Reich works. I agree with that. But without any musical education, nobody could see a similarity between the two.
Fortunately, some smart people have figured out (and proved) that the best features to compute similarity between songs could be found on the internet (and is cultural, indeed). You can obtain precise features to describe a song or an artist by summarizing the words used in amazon reviews, or usenet posts. You can see how similar two songs are by counting their co-occurences in webradio playlists. A Google search will tell you that "Paul Mc Cartney" and "The Beatles" have something in common, because there are approx. 715000 web pages mentioning both names.