Data Augmentation Strategies for Music Composition using Generative Adversarial Networks

Abstract

The field of Algorithmic Art has been following technological advances in Artificial Intelligence and, as Generative Adversarial Networks (GANs) have become popular, applications on art generation began to emerge. For most deep neural networks, large amounts of training data are essential to achieve satisfactory model quality. But there are cases, such as in MIDI musical melodies, where it is not trivial to acquire data in such a high volume. Data augmentation strategies play an important role on these cases. This paper presents a data augmentation pipeline, composed of three strategies, with the objective of improving the quality of a GAN- based musical melody generator. The proposed data augmentation pipeline was compared with a non-augmented dataset and a replicated dataset, which had the same size of the augmented dataset, but composed only of replicas. From the statistical tests performed it can be stated that the augmented dataset outperformed the non-augmented dataset and the replicated dataset, when evaluating the Fréchet Inception distance (FID) score.