Automatic Music Composition With BiLSTM And GANs Algorithms

With the rapid development of deep learning in recent years,there are more and more applications of deep learning in automatic composing.Music is entirely subjective,and the result of beauty is entirely determined by people’s feelings,without any objective criteria,so music creation is a huge challenge to computers,which also prompts people to think that fiction and music creation is the last area of AI.Music creation has its own particularity.First of all,music has time sequence.The music of the current moment is related to the music information before or after the current moment.In addition,in real life,musicians create in a non-linear way,that is,they write a part first and then modify it repeatedly.In order to solve the dependence of music time sequence,this paper first proposes a hybrid network G composed of BiLSTM and NN with time series to generate music.In order to simulate the non-linear composing method of musicians,this paper adopts a pseudo-Gibbs sampling method.In order to generate music,D in GANs is used to judge whether the data distribution of the generated music is similar to that of the real music,and then the information obtained by D is fed back to G to fine tune G network,so that D and G can play games continuously,and finally G is used to generate music.The model proposed in this paper is trained on the data set of Bach’s hymns.After training,model G can generate songs with Bach’s style.The advantage of the model is that it has high flexibility and can add some restrictions of music theory knowledge,such as notes,melodies,stress,etc.according to the needs in the process of production.The first chapter of this paper mainly introduces the development background of automatic composing in symbol and audio domain,the main development direction and research status,the process of some neural networks and automatic composing commonly used in automatic composing,and the ideas of this paper.Because the structure of G network is mainly used for the knowledge of BiLSTM and LSTM,the second chapter of this paper mainly introduces the basic structure of LSTM,the process of forward and backward algorithms,and the main content of the transition from LSTM to BiLSTM.Chapter 3 of this paper mainly introduces the principle of GANs and its role in the direction of automatic composing,the algorithm of GANs and the related content of iteration optimization.In the fourth chapter,firstly,the method of processing Bach’s hymn data is introduced.Then,the network structure of the generating model G based on BiLSTM and NN is proposed,and the function of each part of the network structure and the arithmetic of the generating process are introduced.Finally,the network structure of the discriminant model D is proposed,and the function of the discriminant model and the comprehensive GANS model are introduced.Finally,numerical experiments are carried out on the data set of Bach’s hymns,and the experimental results are shown in Chapter 5.In the form of a questionnaire survey,the students majoring in piano at Northeast Normal University are asked to evaluate the experimental results.