Music generation with generative adversarial network (GAN)

• Main Author: Tan, Yi Zhuang
• Other Authors: Alexei Sourin
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/148183
• Institution: Nanyang Technological University

Music generation using deep learning has recently been gaining quite a bit of traction. Deep learning involves having a neural structure extract features from the dataset to learn any patterns or structures that are involved in the dataset. Most of the starting approaches to generating music, involves using a recurrent neural network unit, the Long Short Term Memory (LSTM). This was where music generation using deep learning started gaining more attention, and more methods was experimented upon and used. The most recent developments have been using Generative Adversarial Networks [1] (GAN) for music generation. It involves 2 “players”, one being the generator and the other being the discriminator. The discriminator would be trained to recognize real and fake or generated data, while the generator would be trained to try and deceive the discriminator by using converting the noise inputs to generated notes or data. This report will look into generating music using GAN, and how more elements of music was generated with a multi-input and output GAN, while maintaining a simplistic form of representation to facilitate understanding and usage. The generated music was then put through a user study to evaluate the effectiveness of the model to generate more complex music using GAN, while maintaining a simpler approach.