Music style migration based on generative Adversarial Networks

Music style migration, a key technique in the field of music generation, aims to convert audio from one style to another. However, existing methods often struggle to balance content preservation, style consistency, and sound quality in many-to-many style migration tasks, and are especially deficient in high-frequency detail preservation and complex style mapping. To address these issues, we propose the CAUNet model for efficient many-to-many music style migration by combining the autoencoder, CycleGAN and U-Net augmented generator modules. The autoencoder module extracts the core content features of the audio, the CycleGAN module ensures the authenticity of the style transformation, and the U-Net augmented generator module improves the detail reproduction through multi-scale feature fusion, so that the generated audio meets the target requirements in terms of both style and sound quality. The experimental results show that CAUNet significantly outperforms the mainstream comparison models in terms of content consistency, style matching, and sound quality, and maintains high generation quality in the complex style combination task. The model provides a reliable framework for realizing many-to-many music style migration and has wide application potential.