Tonal-Atonal Classification of Music Audio Using Diffusion Maps.

In this paper we look at the problem of classifying music audio as tonal or atonal by learning a low-dimensional structure representing tonal relationships among keys. We use a training set composed of tonal pieces which includes all major and minor keys. A kernel eigenmap based method is used for structure learning and discovery. Specifically, a Diffusion Maps (DM) framework is used and its parameter tuning is discussed. Since these methods do not scale well with increasing data size, it becomes infeasible to use these methods in online applications. In order to facilitate on-line classification an outof-sample extension to the DM framework is given. The learned structure of tonal relationships is presented and a simple scheme for classification of tonal-atonal pieces is proposed. Evaluation results show that the method is able to perform at an accuracy above 90% with the current data set.