Training Sequence Generative Adversarial Nets to Compose Music in the abc-notation

Abstract

Creating new music is traditionally seen as a task performed only by humans. Getting computers to compose music would have advantages such as the ability to create more music in the same amount of time. With recent developments in the field of machine learning, music composition by computers seems closer than ever. In this study we have applied the Sequence Generative Adversarial Nets technique as proposed by Yu et al. to the task of generating tunes in the abc notation, a text-based music notation system. The training set we have used was built by Sturm et al. and consists of 23,636 abc songs. We succeeded in training a model that generates valid abc files. These songs were empirically evaluated with the help of 47 participants. No significant difference between the ratings of real tunes and generated tunes was found. However, we think it is possible that even the real tunes were not well-received by the participants which would make this conclusion less valuable.