Generation of Musical Scores from Chord Sequences Using Neurodynamic Model

Abstract

Music generation is one of the important and challenging tasks in the field of music information processing. In this paper, we propose a method to generate music note sequences from a chord sequence based on the dynamics relation learned using a recurrent neural network model. For the dynamics learning model, Multiple Timescale Recurrent Neural Network (MTRNN) is used. The model comprises a hierarchical structure to learn different levels of information. The note sequence and chord sequence dynamics are self-organized into the model based on the similarity of the dynamics. The proposed method inputs a chord sequence into the trained MTRNN to calculate the latent parameter that represents the sequence. The parameter is used to generate the music note sequence using the forward calculation of MTRNN. Experiments were conducted using three music pieces arranged into three variations. The results of the experiment generating music note sequence from chord sequence for trained pieces showed the effectiveness of the method.