Generation of music pieces using machine learning: long short-term memory neural networks approach

Abstract In this article, we explore the usage of long short-term memory neural network (NN) in generating music pieces and propose an approach to do so. Bach’s musical style has been selected to train the NN to make it able to generate similar music pieces. The proposed approach takes midi files, converting them to song files and then encoding them to be as inputs for the NN. Before inputting the files into the NNs, an augmentation process which augments the file into different keys is performed then the file is fed into the NN for training. The last step is the music generation. The main objective is to provide the NN with an arbitrary note and then the NN starts amending it gradually until producing a good piece of music. Various experiments have been conducted to explore the best values of parameters that can be selected to obtain good music generations. The obtained generated music pieces are accepted in terms of rhythm and harmony; however, some other problems exist such as in certain cases the tone stops or in some other cases getting short melodies that do not change