Research On The Algorithm Of Music Generation Based On Deep Learning

As an art reflecting human emotions and thoughts,music has been deeply involved in people's daily life.But it is not easy to create music works.With the development of artificial intelligence and automation technology,the emergence of computer-based automatic music generation technology meets the needs of non-professional music creation.In recent years,the rapid development of deep learning technology had promoted the research of automatic music generation technology.It also effectively avoided the traditional machine learning method to rely on a large number of music knowledge rules and artificial design features,and achieved better results in complex music generation tasks.Due to the lack of good interpretability of deep neural network,it was very difficult to control the behavior of neural model flexibly.At present,many researches on neuro-musical generation are mainly focused on fully automated generation tasks,but with little attention to interactivity.Existing methods of interactive neural music generation mainly focused on the condition-based autoregressive methods and latent variable methods.Generally,these methods were inflexible control modes,high cost in manual labeling and difficult optimization.By analyzing these methods,the contour control based variational auto-encoder for melody generation was proposed.By unsupervised learning conditional autoregressive task with the melody contour labels automatically inferred from the model,the method realized the modeling of local melody contour features.And it used variational inference to encode latent variables of melody,which made up for the deficiency of the former in global attribute modeling.The method could adjust the local feature preferences of generated melodies.And it also could adjust global features of melodies by changing latent variables.Experiments show that the performance of this method is significantly better than the basic recurrent neural network and variational auto-encoder.And it is easy to train.The local and global features of generated samples could be effectively controlled by the contour labels and latent variables which are more independent.