Machine Composition of Korean Music via Topological Data Analysis and Artificial Neural Network

Abstract

Common AI music composition algorithms train a machine by feeding a set of music pieces. This approach is a blackbox optimization, i.e. the underlying composition algorithm is, in general, unknown to users. In this paper, we present a method of machine composition that teaches a machine the compositional principles embedded in the music using the concept of overlap matrix. In (Tran Mai Lan, Changbom Park & Jae-Hun Jung (2023) Topological data analysis of Korean music in Jeongganbo: a cycle structure, Journal of Mathematics and Music, DOI: 10.1080/17459737.2022.2164626), a type of Korean music called Dodeuri music has been analysed using topological data analysis (TDA). To apply TDA, the music data is first reconstructed as a graph. Through TDA on the constructed graph, a unique set of cycles is found. The overlap matrix lets us visualize how those cycles are interconnected in music. We explain how we use the overlap matrix for machine composition. The overlap matrix is suitable for algorithmic composition and also provides seed music to train an artificial neural network.