Investigation on Automatic Music Generation Using Gan and Lstm Networks

Abstract

In this article, the authors propose a few methodologies for composing music using deep learning algorithms and Long-short term memory (LSTM) neural network and Generative Adversarial Networks (GANs). The LSTM model is created by training with a set of input files from a music library. The trained model then synthesizes music when an arbitrary note is provided. The GAN and other variants are trained using a set of midi file accumulated from the piano dataset. The pre-trained GAN model is then used to generate music similar to piano roll. The quality of the music is calculated by comparing the harmony and few other parameters of the synthesized music with the trained files. The music library is made with a set of midi files and based on the chosen library, a unique model shall be created. For the model creation, the library files are converted into a suitable format and encoded in order to make it compatible with the LSTM network. Though the outcome of this experiment is a continuous music, the harmony and notes can still be improved to solve the discontinuity problem. The outcome of this experiment is evaluated using conventional evaluators and also the aesthetics by human observer.