Machine Learning for Music Genre Classification

Classifying music genres according to their audio features is important for many applications including music recommendation systems, music organization, music categorization and content-based retrieval. This study used unsupervised machine learning algorithms for categorizing music genres. Three analytics were applied using new fusion techniques for music genre classification and included K-Nearest Neighbor, Support Vector Machine and Linear Discriminant Analysis. Features used with these machine learning analytics were BFCC, Mel-Frequency Cepstral Coefficients, GFCC, NGCC, LGCC, chroma, spectral roll-off, spectral centroid, spectral bandwidth, ZCR, RMSE, temp, harmonic, and percussive features. These features were extracted using LibROSA and SPAFE python packages. Features were then used on their own or in combination in machine learning models for classifying music genres. The best models were observed when KNN was applied to six features and showed high accuracy of classification of 93%. This in turn confirmed that right fusion technique, combined with features, provided high accuracy in classifying music genres.