Automatic Composition And Arrangement Research Based On Deep Learning

Music is one of the greatest inventions in human history and has a vital influence on human life.However,composing music needs plenty of professional knowledge and instrument skills.How to generate music automatically has become a hot topic in recent years.Many companies and research institutes have done interesting works in this area.For example,Google Magenta project could generate piano by learning massive piano data with machine learning method.Recently,deep neural networks have been applied in music generation by the end-to-end method,without massive manpower and domain knowledge.However,some critical factors related to the quality of music are not well addressed,such as chord progression and rhythm pattern.Chord progression generally exists in pop music,which could guide melody procession.Thus,it is beneficial to capture chord progression as input for music generation.Besides,pop music has rhythm patterns,which make the music more structural and pause suitably.However,existing studies usually generate music note-by-note and without considering the rhythm pattern.Besides,pop music usually has multiple tracks and instruments to accompany the melody,while they should be in harmony with each other,but none of previous work considered the harmony between multiple tracks.At last,music usually has multiple styles,such as classic,jazz and pop.How to control the musical style is a valuable problem need to be explored.To address these challenges,based on deep learning method,this paper studies the music generation problems,the major work and contributions are as follows:1.For pop music generation,we propose an end-to-end melody and arrangement generation framework(MAFG),including melody generation section and arrangement generation section.In melody generation section,we propose a chord based rhythm and melody cross-generation model(CRMCG)to improve the notes relationship as well as music structure.In the arrangement generation section,we propose a multi-instrument co-arrangement generation model(MICA)to generate multi-track music,considering harmony among different tracks.2.For musical style,we propose a multi-style multi-instrument co-arrangement generation model(MSMICA)to control the musical style of generated music.Specifi-cally,we use MICA as the generator and multi-discriminators of multi-style discrimi-nator and harmony discriminator to keep the style and harmony.3.For the above two works,a large number of experiments were performed on real-world music datasets.The experimental results prove the validity of the CRMCG,MICA and MSMICA in this paper.