Design and Research of Interactive Intelligent Melody Arrangement System through Deep Learning and Artificial Neural Networks

In recent years, the rapid development of deep learning techniques has promoted the research of automatic music generation techniques, effectively avoiding the over-reliance of traditional machine learning methods on music knowledge rules and manually designed features, and achieving better results in complex music generation tasks. However, deep neural networks are very difficult to control the behavior of neural models flexibly due to the lack of good interpretability. Most of the current research on neural music generation has mainly focused on fully automated generation tasks with little attention to interactivity. To address these problems, this paper proposes a variational self-encoder for melody generation based on contour label sequence control. The method achieves modeling of local contour features of melodies by unsupervised learning of conditional autoregressive tasks from melodic contour labels automatically inferred by the model; latent variable encoding of melodies using variational inference compensates for the former's deficiency in modeling global attributes.