MICW: A Multi-Instrument Music Generation Model Based on the Improved Compound Word

Abstract

In this work, we address the task of multi-instrument music generation. Notably, along with the development of artificial neural networks, deep learning has become a leading technique to accelerate the automatic music generation and is featured in many previous papers like MuseGan[1], MusicBert[2], and PopMAG[3]. However, seldom of them implement a well-designed representation of multi-instrumental music, and no model perfectly introduces a prior knowledge of music theory. In this paper, we leverage the Compound Word[4] and R-drop[5] method to work on multi-instrument music generation tasks. Objective and subjective evaluations show that the generated music has cost less training time, and achieved prominent music quality.