Welikala, MD, & Fernando, TGI (2021). Komposer V2: A Hybrid Approach to Intelligent Musical Composition Based on Generative Adversarial Networks with a Variational Autoencoder. In K. Arai, S. Kapoor, & R. Bhatia (Eds.), Proceedings of the Future Technologies Conference (FTC) 2020, Volume 1 (pp. 413–425). Springer International Publishing. https://doi.org/10.1007/978-3-030-63128-4_31
Abstract:
Among numerous art forms that exist in the world, music could be considered as an art form that has a cultural value as well. It could be seen that most of the cultures that exist and existed in the world consider music as a part of their culture and the definition of music varies according to culture and social context. Music can be divided into genres and genres can be further divided into sub-genre. Although the dividing lines and relationships between music genres are often subtle, sometimes it is open to personal interpretation and occasionally controversial. Even though research on musical composition has been carried out with the use of different technologies during the past few years, each research has had its positives and drawbacks. Hence, researchers have been experimenting with new methods of musical composition. The purpose and the aim of this research is to identify aforementioned genre specific features and generate music notes which reflect a particular genre so that we can reproduce music that carry similarities of genre that are extinct as well as the genre that are currently being practiced, by using Generative Adversarial Networks with a Variational Autoencoder. In this study, we were able to successfully generate musical melody notes provided a genre. Further a web-based inference tool that allows us to generate musical melody was also developed as a result of this study.
Keywords:
Generative Adversarial Networks, Variational Autoencoders, Musical composition, Convolutional Neural Networks