Research On Folk Song Composition Based On Markov Model And Neural Network

With the development of artificial intelligence technology,artificial intelligence composition has gradually become a popular research direction.Artificial intelligence composition can bring new creative inspiration for composers,and also allow non-professional music lovers to participate in music composition and enjoy the fun brought by artificial intelligence composition.At present,the research on artificial intelligence composition technology for Western music is relatively active,but there is little research on the automatic composition of C hinese folk music.The purpose of this thesis is to study algorithmic composition and apply deep learning and other artificial intelligence technologies to the composition of Chinese folk music.It has important practical significance for inheriting and developing Chinese national culture.The main research work and innovations of this thesis are as follows:1.In the part of data preprocessing,this thesis proposes an up-down sampling coding method,which solves the problem that melody pitch and duration are extracted respectively as independent training feature,and the rhythm relationship between pitch and duration cannot be characterized in the traditional algorithm of composition research,which leads to the network model can't learn music style better.The rhythm of music plays an important role in the expression of the music style.The up-down sampling coding method helps the network model learn the rhythm style of the music by coding the pitch and duration and then sending it to the network model training,so that the composition network model can generate music with distinctive style.2.In the part of network model construction,this thesis takes typical two-part folk songs as the research object,and proposes to mix markov model,Bi-GRU(Bi-directional Gated Recurrent Unit)and curve fitting algorithm to design the compositin network structure based on the advantages of various algorithms.In the network model design of this thesis,markov model is firstly designed to generate the motivational melody based on the knowledge rules of motivational melody in fo lk song composition,so as to provide the overall initial conditions for the subsequent algorithm composition.Bi-GRU which can extract contextual note sequence information is used to learn the style of MIDI folk song data set independently collected to get a prediction model.The prediction model generates the first part folk song melody by combining the motivational melody.And at the same time,the relationship between the two-part folk song melody is studied based on the first part folk song melody,and the second part melody is modeled by the curve fitting method.Through comparison experiments,the accuracy of the test set in the model proposed in this thesis reaches 88%.3.In the part of music compostion quality evaluation,this thesis proposes two evaluation methods:objective evaluation method and subjective evaluation method.The objective evaluation method is to design a criterion suitable for the evaluation of two-part folk song melody through the analysis of music theory rules,which is used to select the curve fitting function most suitable for the generation of the two-part folk song melody.The subjective evaluation method is to invite a number of audiences from different levels to listern to the music,and then use the fuzzy comprehensive evaluation method combined with entropy weight method to comprehensively score each piece of music.The final evaluation results verify the feasibility and the application value of the compostion algorithm designed in this thesis.