Saturday, February 16, 2008

Music Recommenders

Just wanted to point out a really good post comparing the differences between Last.fm and Pandora (circa January 2006). In a nutshell, both services compile databases used to make decisions about which songs/artists are similar to one another and then use these decisions to create playlists automatically based on some seed (e.g. songs that are somehow like songs from the Beatles).  Last.fm does this mainly by looking at who likes what (data cleverly extracted from things like iTunes via their "scrobbler" as well as through direct user preference statements) while Pandora utilizes manually (by them) annotated musical features of the songs themselves such as "subtle use of vocal harmony" and "mild use of rhythmic syncopation".  


Steve Krause's post is a nice, in-depth look at the differences that hold between these two contrasting approaches to musical knowledge acquisition.  Its a great exploration of professional/intentional/manual versus amateur/extensional/semi-automatic content creation - especially because both services are so good!  The only downside is that the post is now a bit dated.  As he predicted it would, Pandora is definitely picking up and almost certainly applying user-generated data now (its even connected to Facebook) - which makes it harder to segregate the two systems. 

Of course, I'm curious what Pandora might be able to achieve by encouraging and enabling their users to annotate songs for them.  Do you think you could correctly identify the mild use of rhythmic syncopation? I probably couldn't, but it would be cool if they showed me how.

0 comments: