Unsupervised Learning of Local Features for Music Classification

In this work we investigate the applicability of unsupervised feature learning methods to the task of automatic genre prediction of music pieces. More specifically we evaluate a framework that recently has been successfully used to recognize objects in images. We first extract local patches from the time-frequency transformed audio signal, which are then pre-processed and used for unsupervised learning of an overcomplete dictionary of local features. For learning we either use a bootstrapped k-means clustering approach or select features randomly. We further extract feature responses in a convolutional manner and train a linear SVM for classification. We extensively evaluate the approach on the GTZAN dataset, emphasizing the influence of important design choices such as dimensionality reduction, pooling and patch dimension on the classification accuracy. We show that convolutional extraction of local feature responses is crucial to reach high performance. Furthermore we find that using this approach, simple and fast learning techniques such as k-means or randomly selected features are competitive with previously published results which also learn features from audio signals.