Timbral modeling for music artist recognition using i-vectors

Music artist (i.e., singer) recognition is a challenging task in Music Information Retrieval (MIR). The presence of different musical instruments, the diversity of music genres and singing techniques make the retrieval of artist-relevant information from a song difficult. Many authors tried to address this problem by using complex features or hybrid systems. In this paper, we propose new song-level timbre-related features that are built from frame-level MFCCs via so-called i-vectors. We report artist recognition results with multiple classifiers such as K-nearest neighbor, Discriminant Analysis and Naive Bayes using these new features. Our approach yields considerable improvements and outperforms existing methods. We could achieve an 84.31% accuracy using MFCC features on a 20-classes artist recognition task.