A Study on Attribute-Based Taxonomy for Music Information Retrieval

We propose an attribute-based taxonomy approach to providing alternative labels to music. Labels, such as genre, are often used as ground-truth for describing song similarity in music information retrieval (MIR) systems. A consistent labelling scheme is usually a key in determining quality of classifier learning in training and performance in testing of an MIR system. We examine links between conventional genre-based taxonomies and acoustical attributes available in text-based descriptions of songs. We show that the vector representation of each song based on these acoustic attributes enables a framework for unsupervised clustering of songs to produce alternative labels and quantitative measures of similarity between songs. Our experimental results demonstrate that this new set of labels are meaningful and classifiers based on these labels achieve similar or better results than those designed with existing genrebased labels.