Research on automatic music generation with multi-track based on melody constraints

Abstract

Aiming at the problems that the melody generated by the existing multi track music generation model is not pleasant and the harmony between tracks is not enough, a multi track music generation model with TransformerX-SpanBERT sequence conditions is proposed. The module is a Generative Adversarial Network structure. Since the generated music has multiple tracks, first, use TransformerXL as the generator and SpanBERT as the discriminator to generate piano track music, and the other tracks have one generator for each track. Therefore, there are multiple generators and one discriminator. First, piano track music is as input, let each track generator to generate their own music. Then, the common discriminator is used to judge the music quality generated by each track and the harmony between tracks. Finally, the global attention mechanism is used to pay attention to the music characteristics to ensure the overall effect of music. The objective evaluation index of music and the subjective score of listeners are used to verify the model and compare it withTransformer-GAN and MuseGAN Multi-track sequential symbol music, the results show that the music generated by the model in this paper has the advantages of more harmonious tracks and better melody, rhythm and coherence, therefore, it can be applied to practical music creation.