An intelligent music generation based on Variational Autoencoder

In this paper, GAN and VAE are combined with deep learning network to generate intelligent music based on music theory rules, and to explore intelligent music generation algorithm. Different from the traditional algorithmic composition, it is not necessary to manually add complex rules, but trains the initial music set, evaluates and filters the music collection, and ultimately generates music via the RVAE-GAN neural network. The fitness function calculates a weighted sum of a series of features of a piece of music, such as pitch and rhythm distribution, and can also calculate a series of theoretical rules of music theory from the distance between a particular piece of music. Furthermore, we take the music theory and practical experience into account, put forward the expressions and a rule function of rhythm, which are given in a mathematical way. On this basis, the semi-supervised algorithm is used to form the chord structure model, combined with the feature extraction of music, and the intelligent generation music based on GAN confrontation generation network and VAE network combined with music theory rules is proposed and proposed. And mass production has important theoretical and practical significance.