Deep Generative Architectures for Automated Music Composition: Optimizing Neural Structures and Multimodal Inputs for Style-Conscious Melody and Harmony Generation

This study explores the application of deep generative models in the field of intelligent composition, focusing on the impact of network architecture optimization and multimodal input integration on music style fidelity and emotional expression. By comparing and analyzing the performance of LSTM, transformer, and hybrid convolution architectures in the melody generation task, experiments were conducted based on a multimodal dataset containing 8,000 sets of lyric melody pairs and 2,000 sets of emotional samples. Test data show that the transformer architecture performs better in key indicators such as BLEU score and tonal distance, and the training efficiency has been significantly improved. The research confirmed that the multimodal orientation of lyric texts, emotional labels, and visual images can effectively improve the emotional coherence of the generated music. In the blind evaluation, music professionals' preference for the generated multimodal works reached 78%. Through examples of style transfer such as jazz transitions and classical variations, the model's potential in contextual creation was verified. This achievement provides a scalable technical framework for the field of AI-assisted composition and music therapy, promoting the practical application of intelligent music generation.