Deconstructing Mozart: A GAN-Style Approach to Raw Audio Processing and Generation

Abstract

We propose a new system for raw audio processing and generation which combines the success of Generative Adversarial Networks (GANs) and WaveNet, a generative model created by DeepMind. This system will generate piano music using only sample piano audio files that are given to it. The use of GANs together with raw audio, rather than derived features, makes this project unique in the realm of machine learning. Most previous works that compose music through artificial means utilize representations of music, rather than raw audio, due to computational complexity, and previous works that use GANs often solve the problem of image synthesis. Current progress of the fully implemented system with WaveNet and GANs are able to nearly mimic the frequency domain characteristics of the inputted data.