Chinese opera music, as a traditional cultural treasure, carries deep historical heritage and unique artistic charm. In this paper, a Transformer-based melody generation model for opera music, Tr-MTMG, is proposed, which realizes the inheritance and innovation of opera style through multimodal time series analysis. Methodologically, the model consists of three parts: data preprocessing network, learning network and generative network, in which the learning network contains six Encoding layer sub-networks, and the Cross-track attention mechanism is used to interactively learn the time series information between different tracks. The experimental results show that Tr-MTMG generates 128 bars of opera music with 13-19 themes, the rate of empty bars is reduced by 1.429%, the ratio of qualified notes is increased to 96.862%, and the overall quality score of subjective evaluation is 3.73 points. The model effectively solves the deficiencies of traditional music generation methods in long-term structural consistency and style maintenance, and generates opera music with rich melodic variations and good structural coherence, which provides technical support for the digital inheritance of opera music.
