Deep learning in automatic music generation

Abstract

Music is a grouping of musical tones from various frequencies. While artists composed through a deliberate arrangement of different notes, nowadays, A.I. programs learn to automatically generate short music through a machinal sequence of distinct notes. This essay compared the utility and efficiency of traditional machine learning (Regression Model) and deep learning methods (LSTM). This research only focused on instrumental classical music and used the MusicNet collection as the primary dataset. The comprehensive experiments are conducted from these two models, which suggests two results. Firstly, the LSTM model generates melodies that better fit the training styles. Secondly, models are better fitted on single music data than on the entire dataset.