Unsupervised automatic music genre classification

Trabalho apresentado no âmbito do Mestrado em Engenharia Informática, como requisito parcial para obtenção do grau de Mestre em Engenharia Informática

In this study we explore automatic music genre recognition and classification of digital music. Music has always been a reflection of culture di erences and an influence in our society. Today’s digital content development triggered the massive use of digital music. Nowadays,digital music is manually labeled without following a universal taxonomy, thus, the labeling process to audio indexing is prone to errors. A human labeling will always be influenced by culture di erences, education, tastes, etc. Nonetheless, this indexing process is primordial to guarantee a correct organization of huge databases that contain thousands of music titles. In this study, our interest is about music genre organization. We propose a learning and classification methodology for automatic genre classification able to group several music samples based on their characteristics (this is achieved by the proposed learning process) as well as classify a new test music into the previously learned created groups(this is achieved by the proposed classification process). The learning method intends to group the music samples into di erent clusters only based on audio features and without any previous knowledge on the genre of the samples, and therefore it follows an unsupervised methodology. In addition a Model-Based approach is followed to generate clusters as we do not provide any information about the number of genres in the dataset. Features are related with rhythm analysis, timbre, melody, among others. In addition, Mahalanobis distance was used so that the classification method can deal with non-spherical clusters. The proposed learning method achieves a clustering accuracy of 55% when the dataset contains 11 di erent music genres: Blues, Classical, Country, Disco, Fado, Hiphop, Jazz, Metal,Pop, Reggae and Rock. The clustering accuracy improves significantly when the number of genres is reduced; with 4 genres (Classical, Fado, Metal and Reggae), we obtain an accuracy of 100%. As for the classification process, 82% of the submitted music samples were correctly classified.