Research On Intelligent Music Generation Based On Deep Learning And Music Theory Rules

Music is an important way for people to regulate their lives and emotional expressions.As the quality of life continues to improve,people’s demand for music has become stronger and stronger in many fields.New forms of entertainment such as games and TikTok videos are inseparable from the rendering of music,but the high cost of human composing,the long cycle time,and the single style of works have the market to explore new methods for composing.One of the most rapidly developing methods is Artificial Intelligence(AI)music generation.While AI composing has produced good results in single track music generation,the development of multi-track music generation is not so satisfactory.For example,there are few models for multi-track music generation,poor harmony of the music generated by the models,and music that is not well connected within and between tracks and does not conform to the basic rules of composition are numerous,any one of which can seriously affect the artistry and aesthetics of the music.In order to effectively solve the above problems,this thesis conducts a research on multi-track music generation by combining the market demand and the current research status of music composition.The main contributions of this thesis are divided into the following three aspects.(1)To generate high-quality multi-track music pieces,this thesis proposes a Transformer-based Multi-Track Music Generation(Tr-MTMG)network to automatically generate music with three instrument tracks:piano,guitar and bass.The model treats the music generation task as a text-like generation task,describing the music in text form and feeding the text into the network as a word vector.A supervised learning process is formed by using the input sequence of the next moment as the target value,enabling the Tr-MTMG network to learn the dependencies between notes effectively.The experimental results show that the model is able to generate audible,smooth and rhythmic multi-track music pieces.(2)In order to solve the problem that the music generated by the generative model does not conform to the rules of music theory,this thesis proposes a Music Rule MultiTrack Music Generation(R-MTMG)network based on music rules.The model introduces a reward network,which consists of a music theory rule reward mechanism and cross-entropy loss constructed in this paper,that determines the value of the generated notes at each moment and feeds it to the generation network,so that the generation network uses the reward network as a guide to create high-quality musical works.The effectiveness of the model is verified by comparing the experimental results with other multi-track generation models.(3)In order to generate models that cannot effectively learn information within a single track of music and between different tracks,this thesis proposes Hybrid Learning Multi-Track Music Generation(HL-MTMG)based on a hybrid learning block.The model constructs a hybrid learning block capable of learning information both within a single track and between different tracks,which consists of Cross-track attention mechanism and Multihead attention mechanism,and it can effectively learn information of real music and improve the harmony of generated music and the prediction accuracy of the model.In order to obtain music that is more in line with music standards,this chapter also produces a dataset for music understanding and compares it with different publicly available datasets to verify that it has a positive effect on the training of music generation models.Finally,in order to verify the effects of the above three aspects,this thesis proposes relative objective evaluations for different solutions.Based on the ratings provided by professionals and amateurs,it is proved that the model proposed in this thesis is more in line with the regularity of compositions,the aesthetics of listeners,and closer to the effects of human compositions in terms of generating effects.