A hierarchical approach to simulation of the melodic harmonization process

In this paper, a hybrid system (method) capable of generating accompaniment to human-created melodies was proposed with the use of machine learning and theoretical knowledge, as well as the idea of song´s global framework to simulate the real composition process. This paper introduced a new method in which the harmonization process were divided into several hierarchies, chords in every hierarchy were expanded from their father hierarchy and were the highest several ranked ones among candidates in the same hierarchy. The improved bidirectional search algorithm was used to simulate this extending process, while music rules were employed to further choose the most appropriate chord sequence. For performance evaluation, a new distance definition was developed to quantify the distance between generated chords and the original chords. The quantitative performance comparisons between the hybrid approach and the traditional HMM approach were also carried out at the end of this paper to show difference between compositional and perceptual process of human beings psychologically.