A Music Classification Method based on Timbral Features

This paper describes a method for music classification based solely on the audio contents of the music signal. More specifically, the audio signal is converted into a compact symbolic representation that retains timbral characteris tics and accounts for the temporal structure of a music piece. Models that capture the temporal dependencies observed in the symbolic sequences of a set of music pieces are built using a statistical language modeling approach. The proposed method is evaluated on two classification tasks (Music Genre classification and Artist Identification) using publicly available datasets. Finally, a distance measu re between music pieces is derived from the method and examples of playlists generated using this distance are given . The proposed method is compared with two alternative approaches which include the use of Hidden Markov Models and a classification scheme that ignores the temporal structure of the sequences of symbols. In both cases the proposed approach outperforms the alternatives. Techniques for managing audio music databases are essential to deal with the rapid growth of digital music distribution and the increasing size of personal music collections. The Music Information Retrieval (MIR) community is well aware that most of the tasks pertaining to audio database management are based on similarity measures between songs [1‐4]. A measure of similarity can be used for organizing, browsing, visualizing large music collection s. It is a valuable tool for tasks such as mood, genre or artist classification that also can be used in intelligent music rec ommendation and playlist generation systems. The approaches found in the literature can roughly be divided in two categories: methods based on metadata and methods based on the analysis of the audio content of the songs. The methods based on metadata have the disadvantage of relying on manual annotation of the music contents which is an expensive and error prone process. Furthermore, these methods limit the range of songs that can be analyzed since they rely on textual information which may