Four-Part Harmonization: Comparison of a Bayesian Network and a Recurrent Neural Network

Abstract

In this paper, we compare four-part harmonization produced using two different machine learning models: a Bayesian network (BN) and a recurrent neural network (RNN). Four-part harmonization is widely known as a fundamental problem in harmonization, and various methods, especially based on probabilistic models such as a hidden Markov model, a weighted finite-state transducer, and a BN, have been proposed. Recently, a method using an RNN has also been proposed. In this paper, we conducted an experiment on four-part harmonization using the same data with both a BN and RNN and investigated the differences in the results between the models. The results show that these models have different tendencies. For example, the BN’s harmonies have less dissonance but especially the bass melodies are monotonous, while the RNN’s harmonies have more dissonance but especially bass melodies are smoother.