MUSIC AND GENRE CLASSIFICATION OF AUDIO DATA

Abstract

Musical genre probably describes music best. Genre classifies music and help people find what they are looking for. Music catalogues have become huge with EMD. Thus, in this context of large musical databases and with EMD becoming extremely popular in the digital world, genre fonns a crucial metadata for audio content description. Thus, techniques of automatic genre identification of music would be a valuable addition to the development of audio information retrieval systems for music. Genre classification techniques using computers have become particularly useful now since manual classification is not feasible as well as it is biased. An expert system for musical genres is impossible to create, as no theoretical framework is available for most kind of music. However, based on similarities between songs and the features present in them, we can make an attempt to cluster similar musical pieces together. This idea forms the basis of Automatic Genre classification techniques. Supervised learning techniques where genre taxonomies are known prior to classification, were used since they lead to a more meaningful classification. The following project report first introduces the different concepts and techniques of genre classification that were implemented during the course of my dissertation. Most of the work in this field till date has been centered around classification involving tinibral features. Possibilities for improvement in that direction have been explored. A special section is devoted to the high-level features and their use in genre classification. Different techniques such as GMMs, KNNs, Genetic Algorithms etc are discussed in details. Next, new techniques such as Novelty Detection are introduced and the classification results are discussed. Several new features are added in the algorithm for genre classification, including a pre-processing step for music/speech classification. The scope of the real-time implementation of the work is also presented.