| Automatic composing technology developing rapidly,with the advent of machine learning techniques in recent years,especially with the wide application of music gradually become a new research hotspot.The existing automatic composition methods are mainly algorithmic composition,and the main realization process can be divided into two parts: extraction of music features and design of algorithm model.Based on statistical knowledge,this paper formulated an improvement scheme on the basis of existing algorithmic composition methods,and proposed an algorithmic composition model based on Bi-LSTM and Self-attention mechanism.The model was applied to empirical analysis,and the experimental results were analyzed to obtain the composition effect of the model.In terms of music feature extraction,based on the characteristics of MIDI files and contour algorithm,this paper uses the Skip-Gram model in Word2 vec to extract music features,so as to obtain the feature vectors of melody notes efficiently and accurately.When designing the composition model,Bi-LSTM model is used for training,which is more suitable for musical note feature sequence,and Self-attention mechanism is superimposed to make the model composition better.Through empirical analysis,it is found that when Bi-LSTM is used as the composition model,compared with one-hot method,the accuracy of musical notes is improved from 46.09% to 88.71% when skip-Gram model is used to extract musical features,and the improvement effect is significant.Meanwhile,on the basis of extracting music features with Skip-Gram,Bi-LSTM superimposed Self-attention mechanism was adopted,and the accuracy of musical notes increased from 88.71% to 91.37%.The results show that the proposed composition method is feasible,and verify the feasibility of the improved idea and the correctness of the improved method. |
