Scholarlink

<p>Despite extensive research in music information processing and retrieval, particularly in the context of Western music, there is a notable lack of focus on the processing and retrieval of Persian traditional music. This study stands as the first to examine the effectiveness of self-supervised pre-trained models across various tasks related to Persian traditional music retrieval. Nava dataset, which is a comprehensive collection of Persian traditional musical solos, is utilized for instrument classification, <i>Dastgah</i> recognition, and artist identification. Three pre-trained models, Music2vec, MusicHuBERT, and MERT, are applied and compared across the three tasks. The results indicate that MERT outperforms the other two models across all tasks. Furthermore, the study explores the music signal duration and model fusion on the performance of each task. The findings demonstrate that using combined representations at different levels of layers and model fusion techniques can remarkably improve accuracy. Finally, fine-tuning the pre-trained models results in further enhancements, achieving state-of-the-art accuracy rates of 99.64% for instrument classification, 24.70% for <i>Dastgah</i> recognition, and 79.25% for artist identification on the Nava dataset.</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

On the effectiveness of self-supervised pre-trained models for Persian traditional music information retrieval

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月3日

查看全文
  • Bagher BabaAli,
  • Pouya Mohseni

摘要

Despite extensive research in music information processing and retrieval, particularly in the context of Western music, there is a notable lack of focus on the processing and retrieval of Persian traditional music. This study stands as the first to examine the effectiveness of self-supervised pre-trained models across various tasks related to Persian traditional music retrieval. Nava dataset, which is a comprehensive collection of Persian traditional musical solos, is utilized for instrument classification, Dastgah recognition, and artist identification. Three pre-trained models, Music2vec, MusicHuBERT, and MERT, are applied and compared across the three tasks. The results indicate that MERT outperforms the other two models across all tasks. Furthermore, the study explores the music signal duration and model fusion on the performance of each task. The findings demonstrate that using combined representations at different levels of layers and model fusion techniques can remarkably improve accuracy. Finally, fine-tuning the pre-trained models results in further enhancements, achieving state-of-the-art accuracy rates of 99.64% for instrument classification, 24.70% for Dastgah recognition, and 79.25% for artist identification on the Nava dataset.