A Symbol-level Melody Completion Based on a Convolutional Neural Network with Generative Adversarial Learning

Abstract

In this paper, we deal with melody completion, a technique which smoothly completes partially-masked melodies. Melody completion can be used to help people compose or arrange pieces of music in several ways, such as editing existing melodies or connecting two other melodies. In recent years, various methods have been proposed for realizing high-quality completion via neural networks. Therefore, in this research, we examine a method of melody completion based on an image completion network. We represent melodies as images and train a completion network to complete those images. The completion network consists of convolution layers and is trained in the framework of generative adversarial networks. We also consider chord progression from musical pieces as conditions. From the experimental result, it was confirmed that the network could generate original melody as a completion result and the quality of the generated melody was not significantly worse than the result of a simple example-based melody completion method.

Variable-Length Music Score Infilling via XLNet and Musically Specialized Positional Encoding

“…Coconet [6] trains a convolutional neural network (CNN) to complete partial music score and uses blocked Gibbs sampling as an analogue to rewriting. Nakamura et al [7] use CNN with deconvolutional layers at the end. The whole model is trained under the framework of generative adversarial networks.…”

“…As illustrated in Figure 3, 6 we describe different attributes of a musical note through six different tokens-three note-related ones, PITCH, DURATION, and VELOCITY, and three metricrelated ones, TEMPO, BAR, and SUB-BEAT. 7 Table 1 shows the vocabulary of the adopted token representation.…”

“…A growing body of research work has adopted deep learning techniques to generate music sequentially, taking only the past context as a condition while generating. This paper deals with a different setting where both the past and future context are given, called music score infilling [1][2][3][4][5][6][7][8][9]. As depicted in Figure 1, the goal of this task is to generate a short piece of symbolic music that fits in the middle gap.…”

“…For simplicity, we use monophonic pieces as examples in all the figures in the paper, though we actually use polyphonic pieces in experiment 7. Even though it is not that reasonable to associate a TEMPO event with each note, we found that the models learn to predict similar tempos for adjacent notes, thus barely lower the quality of music.…”

Generating Melodies from Melodic Outlines Towards an Improvisation Support System for Non-musicians

“…CNNs are a promising approach for creating melodies. In fact, many researchers have used CNNs to create melodies [9,10,11]. A CNN can hierarchically reduce the temporal information in a melody by stacking convolution layers.…”