Music Genre Classification Using Frequency Domain Features

Automatic music genre classification is an active and most popular area of research in Music Information Retrieval (MIR) domain. Such classification enables organized storage and retrieval of music information from a large collection. In this work, we have utilized spectral features that represents important musical properties like timbre, tonality and pitch which play important role in discriminating the genres. For classification, support vector machine and random forest are considered. Experiment is carried out on three different benchmark dataset. Result shows that presented features perform satisfactorily for both the classifiers. Comparison with reported systems reflect the superiority of proposed system. Selection of the feature set is well justified as improved performance is achieved across the classifier and dataset.

I. Introduction

Retrieval of music information has become an active area of research and has gained impetus due to the development of on-line music stores. A music database often contains millions of music tracks. Handling such a huge database and retrieving the desired piece is very difficult. Proper categorization of music data will be of help in the context of proper archival and retrieval. Attributes like genre, singer, emotion are quite important in categorizing the dataset. Music genre recognizes a music track as a representative of a tradition or shared traditions. Manual labeling the genres of a large dataset is quite tedious. Moreover, music genres do not have accurate definition and boundaries because it has factors like cultural, historical, geographical, instrumental and rhythmic structures. Hence it is very subjective and requires expertise for correct labeling. As it has no universal taxonomy, the labeling is very much susceptible to error. So an automatic genre identification system is highly required.