A deep learning-based collaborative generative model for affective and cultural symbols in Xiangju opera music

Abstract

Xiangju opera, a national intangible cultural heritage, encodes rich emotions and cultural symbolism through its musical language. However, computational frameworks that effectively capture the intrinsic coupling between affective expressions and cultural semantics remain scarce. This paper presents a deep learning-based collaborative generative model comprising three hierarchical modules: multimodal acoustic feature extraction, affective-cultural joint representation learning, and Transformer-based conditional music generation. The proposed system is designed with consideration of real-time multimedia communication scenarios, featuring a lightweight architecture suitable for edge deployment. Extensive experiments conducted on a purpose-built Xiangju dataset demonstrate that the proposed model achieves an affective classification accuracy of 91.7% while generating stylistically authentic musical excerpts. With a total parameter count of 42.3M and an inference latency of 124 ms per 5-second segment on a single NVIDIA A100 GPU, the model offers a feasible computational pathway for the digital preservation and intelligent transmission of traditional Chinese opera over modern communication networks.