Neural Networks And Ensemble Based Architectures To Automatic Musical Harmonization: A Performance Comparison

Abstract

Harmony can be defined in a musical way as art that combines several musical notes reproduced simultaneously to create sounds that are coherent to human ears and serve as accompaniment and filling. However, working out harmony is not a simple task. It requires knowledge, experience, and an intense study of music theory, which takes time to reach good skills. Thus, systems capable of automatically harmonizing melodies are beneficial for experienced and novice musicians. In this paper, a comparative study between distinct architectures and ensembles of Artificial Neural Networks was proposed to solve the problem of musical harmonization, seeking consistent results with rules of music theory: Multilayer Perceptron (MLP), Radial Basis Function network (RBF), Echo State Network (ESN), Extreme Learning Machines (ELM), and Long Short-Term Memory (LSTM). For this, a processed and defined melody with symbolic musical data serves as input to the system, having been trained from a musical database that contains melody and harmony. The output is the chord sequence to be applied to the melody. The results were analyzed with quantitative measures and the ability to melody adaptation. The performances were favorable to the MLP, which could generate harmonies according to the objectives.