CPS: Full-Song and Style-Conditioned Music Generation with Linear Transformer
Weipeng Wang, Xiaobing Li, Cong Jin, Lu Di, Qingwen Zhou, Yun Tie
Abstract
Many deep music generation algorithms have recently been able to produce good-sounding music, but there have been few studies on controlled generation. In this process, the human sense of participation is usually very weak, and it is difficult to integrate one’s own musical motivation into the creation. In this study, we will introduce CPS (Compound word with style), a model that can specify a target style and generate a complete musical composition from scratch. We first added the genre meta-information to the music representation and distinguished it from other low-level music representations, thus strengthening the influence of the control signal. We modeled with the linear transformer, while used an adaptive strategy with different settings for different types of music tokens to reduce the probability of disharmonic music. The experiments show that, when compared to the baseline model, our model performs better in terms of basic music metrics as well as metrics for evaluating controlled ability.