Generating Retro Video Game Music Using Deep Learning Techniques

Abstract

Music generation using deep learning is a widely studied field. This thesis focuses on music generation in a constrained and novel environment; retro video game music. The constraints imposed by the environment creates many unique challenges for the generation of musical compositions. In addition, the dataset consists of multi-instrument music, which is rarely studied due to its complexity. An extension to an existing architecture; the Biaxial RNN is presented in order to extend its capabilities to allow for generating multi-instrument arrangements. The resulting implementation is somewhat successful at fulfilling the proposed solution, although one component could not be implemented within the time limit. The result is not pleasant music, but it does give a view into the complex process of multi-instrumental music generation.