Long-term symbolic music generation based on dilated convolutional adversarial networks and LM-VGAN

The rapid development of intelligent algorithms in the field of music creation has promoted the continuous improvement of symbolic long-term music generation technology. However, existing methods still have shortcomings in terms of poor rhythmic continuity and weak style transfer capabilities. Therefore, an approach based on dilated convolutional adversarial networks and long-term music variational GAN is proposed. This method introduces a hierarchical dilated convolution structure to enhance the global receptive field, and combines multi-scale gating with a dynamic memory weighting module to improve historical information utilization and feature selection capabilities. It also uses a Wasserstein generative adversarial network with gradient penalty to stabilize the training process. The outcomes indicated that the research method achieved a score of 0.82 ± 0.01 in note continuity and reduced the structural dynamic time alignment distance to 3.91 ± 0.12, performing significantly better than models such as Transformer. In the style transfer experiment, the style accuracy rate reached 92.34% ± 0.72%, the average opinion score was 4.68 ± 0.04, and the training data requirement was only 42.51 ± 1.20 h. In the diversity test, the lowest information divergence was 1.42 ± 0.08 under latent space dimension 128, and the emotional consistency reached 93.05% ± 0.85%. Under the dynamic memory mechanism, the historical feature utilization rate reached 80.20% ± 0.72%, and the minimum convergence step was 420 ± 4. In summary, this research method enhances the coherence, diversity, and emotional expression capabilities of long-term music generation by optimizing its structure and designing multiple modules that collaborate with each other. This demonstrates its wide range of benefits and prospective uses in the creation of symbolic music.