Research On Generation Of Traditional Chinese Pentatonic Scale Music Based On Three-dimensional Sequential Convolutional Network

Algorithmic music is not only well-known to the public because of its fun,but also shows great potential in the industry.Research on music computing,music information retrieval and music algorithm generation is increasing worldwide.Aiming at the balance between mathematical logic and aesthetic value in the field of music generation,this paper proposes a research method based on three dimensional sequential convolutional neural network.This method is based on the self-established Chinese traditional national symbol music data set,combined with clustering algorithms and deep learning related algorithms,to construct a three-dimensional sequence convolution generation model of traditional Chinese folk music,and the training generates a model that considers both the overall timing creativity and local music semantics.and qualitatively and quantitatively evaluate the generation effect of traditional Chinese pentatonic music.The main contents of this article are as follows:(1)Data collection,processing and analysis.In this paper,the music data based on MIDI format is firstly captured from the network based on crawler,and the format of data is preprocessed by format analysis,data alignment and modulation,so as to establish the Chinese traditional ethnic symbol music data set(TCPS).Then,text clustering method was introduced to preprocess TCPS music data set,such as feature extraction,and three clustering methods,K-Means,OPTICS,Birch,based on division,density and hierarchy,were used to analyze TCPS Chinese traditional folk music data set.Finally,the nationality of TCPS music data set is analyzed and verified with the theory of Traditional Chinese folk music.(2)Construct three-dimensional sequential model of folk music.Based on BiLSTM framework,a new 3D sequential neural network model(3D-SCN)is described,which combines the translation invariance of convolutional neural network to output the joint distribution of note sequences and realize the tone invariance of note sequences.The TCPS data set was used for training,and the parameters were adjusted and compared with RNN,GRU,LSTM,BiLSTM,3D-RNN,3D-GRU,TP-LSTM-NADE models for quantitative evaluation.The 3D-SCN model showed the optimal effect and the prediction accuracy reached 99.04%.(3)Construct three-dimensional sequential convolution model of folk music.On the basis of 3D-SCN,a three-dimensional sequential convolution neural network(BoYaTCN)based on hierarchical attention mechanism was proposed,and BoYaTCN was applied to the Chinese traditional folk music data set for training.After parameter tuning and ablation experiment,CNN,Residual CNN,DenseNet,DP-TCN,MH-TCN and BoYa+TCN models were used to quantitatively evaluate the substantial impact of each algorithm module on the model.BoYaTCN model showed the optimal effect and the prediction accuracy reached 99.12%.(4)Evaluation of folk music generation results.In addition to evaluation by objective machine learning indicators,this paper also conducts subjective user evaluation on 403 people from five perspectives of totality,harmony,rhythm,structure and nationality,including 86 musician users(practitioners or researchers in the music industry)and 317 non-musician users.Qualitative analysis of music generation results is also carried out through visualization of a large number of music data.The model in this paper has been proved to be able to produce pentatonic folk music with beautiful melody,harmonious coherence and distinctive Chinese traditional pentatonic characteristics,which also conforms to certain musical grammatical features.At the same time,this model can also be used for other artistic creation that needs to consider time series and local integrity.