The Algorithm For Generating Background Music Based On An Improved Multi-Track Sequential Generative Adversarial Networks

Music generation is an emerging field in artificial intelligence.As society develops,people’s pursuit of art and culture is increasing,and there is a growing demand for original music in various industries.This article proposes and implements an algorithm for generating background music based on an improved Muse GAN model.The main contents are as follows:(1)In order to avoid problems such as disjointed musical phrases and fragmented notes in the generated music,this article further processes the existing dataset.To make the transition between each section of the music smooth and natural,the concept of frame-by-frame processing,which is a common method in speech signal processing,is introduced to ensure the continuity of the music.(2)For the demand of generating background music for videos,this article associates the time,motion speed,and motion saliency in the video with the beat,note density,and note intensity in the music.Based on the advantages and disadvantages of the Muse GAN model,several improvement measures are proposed in this chapter: changing the generator model to a model with a cyclic structure;mathematically modeling the music theory and setting different reward and punishment values based on the importance of the theoretical content;combining the advantages of the two temporal models of Muse GAN to propose a new temporal model;adding a feature extractor to the front end of the generator to extract the information features of the generated music samples and input them into the next round of training.In order to reduce the impact of fragmented notes,the operation of average pooling layer is used to smooth the result samples.This measure aims to improve the coherence and stability of the notes.(3)The results of the model are compared through subjective evaluation and objective evaluation indicators.In the subjective evaluation,participants are divided into three categories: professional musicians,music enthusiasts,and general audiences,and the music samples generated by the model are evaluated.The results show that the improved model is favored by most evaluators in the roles of professional musicians and music enthusiasts,while there is little difference from the Muse GAN model among general audiences.It should be noted that there may be some sampling errors in the statistical sample analysis.After completing the subjective evaluation,the same objective evaluation indicators as the Muse GAN model are used to evaluate the samples.The results show that the improved model performs better in multiple objective indicators.The research in this paper proves the feasibility of using the improved Muse GAN model to generate background music and provides a new solution for the current music generation field.