Research On Deep Learning Automatic Composition Based On MIDI Music

In recent years,the development of music has shown a trend of popularization.The professionalism of artificial composition is too high to meet people's growing music appreciation needs.At this time,a composition method that can efficiently and quickly generate music is needed,so the automatic composition method came into being.The method of automatic composition can usually be divided into two key parts: the representation of music features and the design of composition model.This thesis gives optimization schemes for the MIDI music files from the above two parts to complete the research of automatic composition work.The main research work and innovations of this thesis are as follows:(1)For the extraction and representation of music features,this thesis draws on the word vector model for semantic analysis of natural language processing and proposes a method for generating note feature vectors based on contextual semantic encoding.It solves the problem that the traditional automatic composition uses one-hot encoding to express the music features and cannot describe the contextual semantic information of the music features.The note feature vectors generated in this thesis are no longer independent and orthogonal,which can well represent the contextual semantic information of music features and is an optimized solution for the representation part of music features.Through comparative experiments,this thesis shows that the use of the note feature vector proposed in this thesis to represent music features and automatic composition work can effectively improve the accuracy of note prediction.(2)In the design part of the automatic composition model,this thesis proposes an automatic composition model based on Bi-GRU(Bi-directional Gated Recurrent Unit)network and self-attention mechanism,which solves the problem that the traditional automatic composition model is difficult to study deeply the timing information of music and flexible dependency information due to the network structure.The Bi-directional Gated Recurrent Unit network can not only learn the forward dependence information of music,but also learn the backward dependence information of music due to its bidirectionality,which better characterizes the timing information of music.The self-attention mechanism flexibly adjusts the dependency relationship between music features by assigning different weights and highlights key information in the process of note prediction.It makes up for the limitation of the recurrent neural network in exploring the dependent information due to the short-term memory that makes the connection between the discovered notes weaker and weaker.Therefore,the self-attention mechanism optimizes the automatic composition model.(3)In the part of automatic composition evaluation,this thesis proposes an objective and subjective evaluation method.Among them,the objective evaluation is on the criterion of the note prediction accuracy of the automatic composition model.The note prediction accuracy of the automatic composition scheme proposed in this thesis reaches 81.93%,90.15% and 92.62% on Top1 Accuracy,Top2 Accuracy and Top3 Accuracy.In the subjective evaluation,develop an online automatic composition evaluation system,invite music enthusiasts to score and perceptually evaluate the effect of the automatic composition work based on their subjective listening experience,The two evaluation methods jointly ensure the accuracy and vividness of the music produced by the automatic composition scheme in this thesis.