Universal Prediction Applied to Stylistic Music Generation

Capturing a style of a particular piece or a composer is not an easy task. Several attempts to use machine learning methods to create models of style have appeared in the literature. These models do not provide an intentional description of some musical theory but rather use statistical techniques to capture regularities that are typical of certain music experience. A standard procedure in this approach is to assume a particular model for the data sequence (such as Markov model). A major difficulty is that a choice of an appropriate model is not evident for music. In this paper, we present a universal prediction algorithm that can be applied to an arbitrary sequence regardless of its model. Operations such as improvisation or assistance to composition can be realised on the resulting representation.