Scholarlink

<p>In recent years, the field of music generation has yielded a range of promising results, especially due to progress in deep learning and diffusion-based modelling. However, most existing models focus predominantly on Western music traditions, with limited attention to Chinese traditional music, especially the guzheng, an instrument rich in cultural imagery and intricate performance techniques. Current research remains insufficient in areas such as imagery modelling, timbral fidelity, and style-specific control. To address this gap, we propose a novel multimodal music generation framework that generates guzheng music conditioned on Chinese classical poetry. Based on a diffusion transformer architecture (DiT), the system integrates a T5 text encoder and a cross-attention mechanism to incorporate the imagery and semantic features of the input poetry into the generation process. To support this work, we constructed a large-scale and high-quality paired data set of Chinese poems and guzheng music and performed comprehensive evaluations of audio quality (FD, KL), cross-modal imagery-semantic alignment (CLAP), and the reproduction of specific guzheng playing techniques. Experimental results demonstrate that our model achieves competitive or superior performance compared to existing systems in objective metrics and receives high scores in subjective evaluations of structural coherence, melodic fluency, and technique fidelity. This study explores a new direction for the use of artificial intelligence in Chinese traditional music, presenting a multimodal creative paradigm ’poetry-to-music’ that not only expands the scope of music generation research, but also contributes a novel technical framework for the preservation, dissemination, and reinterpretation of China’s intangible cultural heritage. The code and data are available at <a href="https://huggingface.co/NMLAB8/GZGEN">https://huggingface.co/NMLAB8/GZGEN</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

AI-driven generation of guzheng music from classical Chinese poetry: toward a new paradigm of creative practice in Chinese traditional Music

  • 期刊论文

  • 发表日期：

    2025年10月13日

查看全文
  • Jiaxiang Zheng,
  • Moxi Cao,
  • Chongbin Zhang

摘要

In recent years, the field of music generation has yielded a range of promising results, especially due to progress in deep learning and diffusion-based modelling. However, most existing models focus predominantly on Western music traditions, with limited attention to Chinese traditional music, especially the guzheng, an instrument rich in cultural imagery and intricate performance techniques. Current research remains insufficient in areas such as imagery modelling, timbral fidelity, and style-specific control. To address this gap, we propose a novel multimodal music generation framework that generates guzheng music conditioned on Chinese classical poetry. Based on a diffusion transformer architecture (DiT), the system integrates a T5 text encoder and a cross-attention mechanism to incorporate the imagery and semantic features of the input poetry into the generation process. To support this work, we constructed a large-scale and high-quality paired data set of Chinese poems and guzheng music and performed comprehensive evaluations of audio quality (FD, KL), cross-modal imagery-semantic alignment (CLAP), and the reproduction of specific guzheng playing techniques. Experimental results demonstrate that our model achieves competitive or superior performance compared to existing systems in objective metrics and receives high scores in subjective evaluations of structural coherence, melodic fluency, and technique fidelity. This study explores a new direction for the use of artificial intelligence in Chinese traditional music, presenting a multimodal creative paradigm ’poetry-to-music’ that not only expands the scope of music generation research, but also contributes a novel technical framework for the preservation, dissemination, and reinterpretation of China’s intangible cultural heritage. The code and data are available at https://huggingface.co/NMLAB8/GZGEN.