| Music is a time-honored auditory art that nourishes every corner of the cultural industry,It is an important support for other cultural industries.Music creation is a time-consuming and labor-intensive process that requires professional knowledge of music theory and related musical skills.With the development of Deep Learning technology,deep neural networks have also been applied to music generation.The music generated based solely on music theory rules is too rigid and lacks diversity.Music generation model based on Deep Learning can learn the relationship between the notes in the samples,and the musical pieces it generated are richer in form.However,most of the current Deep Learning-based music generation models do not consider music theory rules,or only consider some relatively simple music theory elements such as chord progressions,which are easy to generate dissonant harmonies and melody progressions.In addition,most previous models treated notes and rhythm as related factors,but in fact,a song can be arranged into multiple rhythms,so there is no strong correlation between rhythm and notes.At the same time,most of the previous models only support the generation of monophonic melody,and few models support the generation of polyphonic melody.In addition,the previous work seldom considered the harmony between multiple tracks when generating multi-track music.Reinforcement Learning is a method of learning through trial and error,which mainly guides the behavior of the agent through the rewards obtained by interacting with the environment.The music generation algorithm based on Deep Reinforcement Learning can not only learn relevant music knowledge from the samples through deep neural network,but also make the music generated by the model follow certain music theory rules through the reward based on the music theory rules.Therefore,based on Deep Reinforcement Learning,this thesis studies the problems related to music generation,and the major research contents are as follows:1.For the generation of rhythm and melody for pop music,this thesis establishes a rhythm generation network and a melody generation network respectively for the generation of rhythm and melody.In order to support the generation of polyphonic melodies,this thesis introduces a multi-label classification technique to enable the melody generation model to generate multiple notes at the same time.For the problem of music datasets,this thesis designs a data processing module to separate the notes and rhythm data from the collected music data,which are used to train rhythm generation and melody generation network respectively.2.For the generation of multi-track music,this thesis realizes the generation of multi track music by establishing multiple agents based on the AC algorithm in Reinforcement Learning.In order to make the multiple audio tracks relatively harmonious,this thesis build a global reward network to establish connections between multiple agents,and realizes the harmony constraint between tracks through the global reward.3.To validate the single-track music generation model ACMGM and the multi-track music generation model MACMGM proposed in this thesis by experiments. |
