RNN Composition of Thematically Diverse Video Game Melodies

Abstract

Historically, creative action to the extent of composing music was exclusive to humans. However, machine learning is challenging prior paradigms by leading computers to the frontier of human-level intelligence. Integration between computers and the arts began in 1951 when Alan Turing transposed music for a computer to perform (National Computing Laboratory, Manchester, 1951). Recently, considerable progress has been made due to breakthroughs in both the graphics processing unit and deep learning in the early 2000s. Most efforts emphasize melodic and harmonic styles (Hadjeres G, Pachet F, Nielsen F, in Deepbach: a steerable model for bach chorales generation, arXiv preprint arXiv:1612.01010 , 2016; Huang A, Raymond W, in Deep learning for music, arXiv preprint arXiv:1606.04930 , 2016), but there is little progress in thematic variation within a genre. By utilizing Google’s Magenta, I trained a deep recurrent network to model strongly thematic melodies by using a diverse corpus of video game music. This research challenges the capabilities of recurrent neural networks since throughout a game many themes are presented, ranging from war, defeat, and victory, to love and death. The aesthetics of music are ultimately subjective; nevertheless, my analysis, to mitigate sociocultural biases in music, evaluates generated songs on modulation, strength of rhythm, and repetition. This deep sequence-based model, trained on 2933 video game melodies, recognized unique motifs and composed compelling and distinct themes.