Scholarlink

<p>Lyric-to-melody (L2M) generation has garnered significant attention in recent years. However, existing L2M systems face challenges in integrating user-specified and automatically generated lyrics, ensuring structural coherence in melodies, and achieving precise lyric-melody alignment. To alleviate these issues, this paper proposes Chat Structure Transformer (CST), an L2M system that combines ChatGPT with a Structure Transformer. Specifically, CST leverages ChatGPT’s advanced text generation capabilities to ensure thematic consistency and generate corresponding lyrics while also accommodating user-specified lyrics, thereby enhancing lyric generation flexibility. Meanwhile, CST incorporates the Structure Transformer, which introduces the StruAttention module for automatic recognition of musical structures and employs a customized loss function based on reinforcement learning principles. These components collectively enhance the structural coherence and lyric-melody alignment in the generated melodies. Both subjective and objective experimental evaluations demonstrate that CST produces higher-quality melodies compared to previous systems. Our code is available at <a href="https://github.com/liuasdeu/cst">https://github.com/liuasdeu/cst</a>. Our music is available at <a href="https://github.com/liuasdeu/cst/tree/main/evaluation/cst">https://github.com/liuasdeu/cst/tree/main/evaluation/cst</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CST: a melody generation method based on ChatGPT and Structure Transformer

  • 期刊论文

  • 发表日期：

    2025年5月11日

查看全文
  • Ruhan He,
  • Ruixue Liu,
  • Tao Peng,
  • Xinrong Hu

摘要

Lyric-to-melody (L2M) generation has garnered significant attention in recent years. However, existing L2M systems face challenges in integrating user-specified and automatically generated lyrics, ensuring structural coherence in melodies, and achieving precise lyric-melody alignment. To alleviate these issues, this paper proposes Chat Structure Transformer (CST), an L2M system that combines ChatGPT with a Structure Transformer. Specifically, CST leverages ChatGPT’s advanced text generation capabilities to ensure thematic consistency and generate corresponding lyrics while also accommodating user-specified lyrics, thereby enhancing lyric generation flexibility. Meanwhile, CST incorporates the Structure Transformer, which introduces the StruAttention module for automatic recognition of musical structures and employs a customized loss function based on reinforcement learning principles. These components collectively enhance the structural coherence and lyric-melody alignment in the generated melodies. Both subjective and objective experimental evaluations demonstrate that CST produces higher-quality melodies compared to previous systems. Our code is available at https://github.com/liuasdeu/cst. Our music is available at https://github.com/liuasdeu/cst/tree/main/evaluation/cst.