An Analysis of Implementing a GAN to Generate MIDI Music

Music generated with machine learning has the potential to serve as inspiration and accompaniment for all kinds of musicians. Although neural network technology has been applied to music generation before, these applications have often implemented large and complex networks. This research aimed to use a generative adversarial network (GAN) to generate multi-instrumental, complex music while implementing a simpler structure than past models. This required a data pre-processing algorithm to simplify the encoding of MIDI files. A simple GAN was developed in PyTorch, and generator outputs were scaled to values within the bounds of the MIDI Standard. The GAN ultimately produced MIDI data, but the results were not distinctly musical and lacked a clear melody, perhaps because the network lacked sufficient complexity to process or replicate the temporal structure of music.