With the development of music artificial intelligence, the style modeling and generation technology of traditional folk instrumental music has gradually become a research hotspot. As a representative of traditional Chinese plucked instruments, Guzheng is a typical difficulty in music style modeling due to its rich playing techniques, complex rhythmic structure, and varied timbres. In this paper, we propose a TransformerXL network architecture that combines rhythmic drift modeling, style discrimination mechanism and segment-level memory optimization for high-fidelity Guzheng music style generation and interactive applications. In this study, we adopt a music sequence modeling approach based on REMI event representation, and introduce the RSCLN (residual-normalized fusion) mechanism and segment-level recurrence to enhance the model’s ability to model contextual rhythmic, intensity, and structural information over long distances. In addition, we constructed a performer-generative AI-audience interaction system to realize the closed loop of style migration, score generation, audio synthesis and multimodal feedback for Guzheng music. The model in this paper significantly outperforms existing mainstream methods such as LSTM, TransformerXL, and TSD-GAN in several subjective and objective evaluation metrics, and obtains a style similarity rating (SSR) of 92.7% and a comprehensive listening score of 4.45, as well as an excellent performance in terms of rhythmic naturalness and performance appropriateness.
Li et al. (2026) studied this question.