SIGNATURE WORK
CONFERENCE & EXHIBITION 2022

Many attempts have been made to automatically generate music of certain styles using deep learning techniques. Among them, recurrent neural network (RNN) and Long short term memory (LSTM) are suitable for this task because of the capability to model sequential data and learn long-term dependencies. This research aims at first modifying and training a DeepJ model – an RNN model containing LSTM layers to learn from specific musicians’ styles. Then, the method of Reinforcement Learning Tuner (RL Tuner) is used to enhance the performance of this LSTM model. An RL reward function is built to train this model so that it follows certain musical theory rules while still retaining proximity to the prior probability distribution learned from the data. The generated pieces are evaluated by trained one-class classifiers of One-Class Deep SVDD. Innovations include the combination of RNN models and reinforcement learning (RL) to create musical pieces of certain styles and the usage of one-class classifiers for evaluation purposes. Results show that the proposed model produces music with higher consistency, more diversity and fits the musicians’ styles better.