Research on Generating Xi'an Drum Music Based on Generative Adversarial Network

Currently most AI music generator platforms mainly focus on creating modern music, without focusing on the creation of folk music. And most music generation neural network models use recursive neural networks and their variants, resulting in excessively repetitive music information, greatly reducing the fun of music. When using only generative adversarial networks (GANs) to generate music, it is easy to experience unstable training, vanishing gradients, and mode collapse. Failure to consider time dependence can reduce the authenticity of generated music. In order to solve the problem that most of the conventional deep learning GAN music generation network models do not consider music theory rules in music generation and Xi'an drum music faces the scarcity of tracks and difficulties in composition, we proposed a GAN network music generative model based on music theory rules and characteristics of Chinese Folk Songs. When generating music, the generator CNN and the regulator CNN train together, and generate music melodies under the constraints of the voice column CNN module. The generated results are sent to the discriminator CNN, which submits feedback to the generator CNN for competition, to compensate for the shortcomings of previous technology and ultimately generate better melodies, thereby improving the authenticity, sweetness, and rhythm of the generated music.