A Sentiment-Controllable Music Generation System Based on Conditional Variational Autoencoder

Intelligent music generation is a significant application of artificial intelligence. To better achieve the function of video background music generation, finer-grained control is necessary to enable emotional transitions in music. Therefore, distinct from the variational autoencoder model, this study combines convolutional neural networks with conditional variational auto encoders to improve the intelligent music generation model. We propose enhancements such as the measure-level regularization mechanism and Kullback-Leibler cyclic annealing mechanism, which not only optimize the posterior collapse issue but also further improve the model's representation performance in the latent space and its controllability in generation, aiming to achieve better music representation and finer-grained generative control.