Research On Composition Of Symbolic Music Algorithm Based On Deep Neural Network

Composing music requires music literacy,and composing music requires the composer to master the basic music theory,chords,modes,forms and other professional knowledge,and it also takes a lot of time and energy to complete.This not only brings a high learning cost for ordinary people who want to enter the industry,but also brings a high cost of energy for composers,how to use algorithms to reduce this cost has become an important research direction.In the current related research,there is a lack of means to improve the quality of long sequence music generation in algorithmic composition methods,and there are few studies on polyphonic music generation.In order to solve the above problems,this paper focuses on the method of symbolic music generation.The main research contents are as follows:(1)in order to solve the problem that the quality of RNN(Recurrent Neural Network)neural network composition decreases sharply with the increase of the length of generating sequence,the hybrid neural network algorithm is adopted to improve it.The structure of single neural network model is introduced,and the models of RNN,GAN(Generative Adversarial Networks)and VAE(Variational Auto Encoder)are analyzed respectively.The advantages of the three are combined by the hybrid neural network,and a monosyllabic MIDI music generation model RNN-VAE-GAN based on the hybrid neural network is proposed.The variant GRU model of RNN is introduced to bear the minimum node of the model,and the freezen mechanism is added to optimize the gradient disappearance problem in the GAN model.Finally,the comparison experiment of the generation effect is carried out,and the generation effect is compared with other current composition algorithms with hybrid neural network structure.The experimental results show that compared with other contrast algorithms,RNN-VAE-GAN has a certain improvement in the generation effect,and can obtain better melody variability without improving the fragmentation.(2)in view of the fact that the current neural network composition method is difficult to realize polyphonic music generation,a polysyllabic ABC music generation algorithm based on pre-training and transfer learning is proposed.Through the parametric transfer method based on ABC format,the GPT-2(Gererate Pre-Training Model2)model is transformed into music generation model GPT-Music,and two objective evaluation indexes and two subjective evaluation experiments are designed to verify its effectiveness.The experimental results show that GPT-Music can "understand" ABC music grammar better and achieve better music generation effect than other comparison algorithms.In addition,in the music Turing test,it has been difficult for subjects to distinguish between music generated by the model and real music clips.(3)finally,an online music generation system is implemented based on GPT-Music model.Through web access,the system realizes the functions of account registration,login and management,automatic music generation and music continuation,ABC music editing,online playback,staff display and download and so on.According to the performance test results,the performance of the system meets the requirements of the application scenario.