With the continuous development of artificial intelligence technology and deep learning methods in the field of music information processing, style transfer and multi-voice balance control of choral works have gradually become an important issue in computational musicology. This paper aims to construct a style transfer and voice part balance algorithm for choral works based on generative adversarial networks. Based on 312 choral works and 6840 valid sample fragments, this study jointly collects audio and music data, and completes the unified preprocessing through timing alignment, loudness normalization and multi-modal feature coding. Focusing on the characteristics of chorus style expression and voice organization, we extract information such as spectrum envelope, harmony density, rhythm intensity and voice energy ratio, and introduce cycle consistency constraint, content preservation constraint and voice balance regularization term into the bidirectional generative adversarial framework to realize the collaborative modeling of chorus style transfer and hierarchical optimization. Experimental results show that the proposed method achieves 0.87, 91.6%, 0.041 and 4.36 in style similarity, main melody retention rate, voice part balance error and subjective listening score, respectively, which are better than basic CycleGAN, Transformer-GAN and AutoMix-Net. The results show that the algorithm has good application value in chorus intelligent composition and digital music processing.
