Symbolic Domain Music Generation System Based on LSTM Architecture

Abstract

With the rapid advancement in artificial intelligence, the deep learning technique has changed its application domain from mere classification task to more of generative task. As music is an art of systematizing sounds, this sequential nature of music can be utilized by deep learning techniques like Long-Short Term Memory (LSTM) to generate music. This work emphasizes on LSTM based neural network architecture for generation of music and evaluating how we perceive the generated music. The model is trained on a Classical Piano dataset and implemented using Keras deep learning library. The Mean Opinion Score is used to evaluate the quality of the music generated by the proposed model. The result shows that the proposed model was able to create listenable compositions adhering to human aesthetics.