Automatic music generation by deep learning

Abstract

This paper presents a model capable of generating and completing musical compositions automatically. The model is based on generative learning paradigms of machine learning and deep learning, such as recurrent neural networks. Related works consider music as a text of a natural language, requiring the network to learn the syntax of the sheet music completely and the dependencies among symbols. This involves a very intense training and may produce overfitting in many cases. This paper contributes with a data preprocessing that eliminates the most complex dependencies allowing the musical content to be abstracted from the syntax. Moreover, a web application based on the trained models is presented. The tool allows inexperienced users to generate automatic music from scratch or from a given fragment of sheet music.