Activity: Talks and presentations › Talks and presentations in private or public companies
During the past decade, many researchers have tackled the problem of making computers automatically recognize the genre of recorded music. This is an important problem because it can, among other things, ameliorate the deluge into large archives unlabeled, mislabeled, but always poorly labeled, recorded music. Work in this area in 2001 achieves a mean accuracy of 61% in ten different genres in the GTZAN dataset. Another work from 2006 reaches 83% mean accuracy for GTZAN. And work from 2009 and 2010 claims to observe 91% mean accuracy for GTZAN. With genre so difficult to define, and seemingly based on factors more broad than sound, these are remarkable results. In this talk, I argue from results of three simple experiments that the improvements we see are an unfortunate consequence of excellent discrimination based on confounding factors having little to do with music genre.