An Algorithmic Composition System Based On Neural Networks

Music is one pursuit and habit of human beings by nature.As Long before,philosophers have been interested in how human beings are appreciating music,and how human beings are composing music.Long ago human beings had a series of theories of music.And in the history of music progressing,composers have never stopped the summaries of methods in composition.With the development of modernization,and with the proposition of artificial intelligence,human beings are wondering whether computers can compose music by themselves.And so the research of algorithmic composition comes.The research of algorithmic composition can not only be used in computer aided innovation application,but also mirror the process of perceive and innovation of ourselves.This paper is based on the thought and technology of artificial intelligence,and do some elementary works in algorithmic composition.First,we summarize some common characteristics of music which could be used in learning,and how we represent music data.Second,we introduce the overall procedure of algorithmic composition by recurrent neural networks,and the problem of gradient disappear which we will face.Long Short-term Memory could avoid this problem.This paper also propose clustering music data could benefit following learning process,and produce significant data distribution,which is rare in related research.Last,based on several knowledge introduced in this paper,we propose a music composition system.This system could learn by the data set reserved,or by data set specified.User could also build a data set,in which every music data is similar to some music specified.After system learning,user could specify input vector and play length,and system will play music composed by itself.At the end of this paper,we summary some drawbacks of our work,and how we could improve our work.