Automatic Music Annotation

In the last ten years, computer-based systems have been developed to automatically classify music according to a high-level musical concept such as genre or instrumentation. These automatic music annotation systems are useful for the storage and retrieval of music from a large database of musical content. In general, a system begins by extracting features for each song. The labels and features for a set of labeled songs are used by a supervised learning algorithm to produce a classifier. This classifier can then be used to provide labels for unlabeled songs. In this paper, we examine commercial and academic approaches to musical annotation involving genre, instrumentation, rhythmic style, and emotion. We also describe various musical feature extraction techniques that have been developed for musical genre classification systems. Lastly, we suggest the use of latent variable models as an alternative to the supervised learning approach for music annotation. We describe the correspondence latent Dirichlet allocation (Corr-LDA) model, which has been used for image annotation, and discuss how this model might be adapted for music annotation.