Generation of Two-Voice Imitative Counterpoint from Statistical Models

Generating new music based on rules of counterpoint has been deeply studied in music informatics. In this article, we try to go further, exploring a method for generating new music based on the style of Palestrina, based on combining statistical generation and pattern discovery. A template piece is used for pattern discovery, and the patterns are selected and organized according to a probabilistic distribution, using horizontal viewpoints to describe melodic properties of events. Once the template is covered with patterns, two-voice counterpoint in a florid style is generated into those patterns using a first-order Markov model. The template method solves the problem of coherence and imitation never addressed before in previous research in counterpoint music generation. For constructing the Markov model, vertical slices of pitch and rhythm are compiled over a large corpus of dyads from Palestrina masses. The template enforces different restrictions that filter the possible paths through the generation process. A double backtracking algorithm is implemented to handle cases where no solutions are found at some point within a generation path. Results are evaluated by both information content and listener evaluation, and the paper concludes with a proposed relationship between musical quality and information content. Part of this research has been presented at SMC 2016 in Hamburg, Germany.