Deep Learning-Based Music Generation

Abstract

Music is one of the greatest inventions in human history. Traditionally, music composition is time-consuming and complex, requiring master sound knowledge based on music theory and musical intuition. In recent decades, deep learning have been applied in music generation, and it has experienced the process from simple sequence generation to multi-track generation considering musicality, multiple methods are implied in study to generate better music and combine existing music theory with deep learning technology, while the current technology already allow people composite easily even without domain knowledge and massive manpower. This paper offers an overview of automatic music generation task, covering majority of the currently popular deep learning-based music generation models. In addition, in latter section discusses how to unify objective criteria for music on subjective way, proposes existing deficiencies and expands possible directions. The research in this review has significant foreshadowing meaning and reference value for developing music generation in the future.