Scholarlink

<p>The subjective evaluation of vowel elongation (madd) in Qur'anic recitation presents a challenge for standardized assessment. This study proposes a two-stage hybrid convolutional recurrent neural network (CRNN) framework to detect and classify madd occurrences in the recitation of Surah Al-Fatihah. The model is trained on a curated dataset of 480 audio recordings, which are systematically segmented into 8,640 labeled madd instances, comprising 4,800 madd Thabi’i, 3,360 madd ‘Aridh Lissukun, and 480 madd Lazim Mutsaqqal Kilmi. The study investigates eight acoustic features, including mel-frequency cepstral coefficients (MFCCs), formants, and onset energy. An ablation study identifies MFCC as the most critical feature, as its removal reduces classification accuracy to 44.9%, while other features provide complementary information for spectral and vocal tract representation. The proposed hierarchical framework first distinguishes between madd Thabi’i and madd Far’i, followed by fine-grained classification, effectively addressing the limitations observed in single-stage models. Experimental results show that the model achieves 98% accuracy on the test dataset and maintains strong generalization performance on unseen real-world recordings with an overall accuracy of 98.33%. These findings demonstrate that the hybrid CRNN effectively captures both spectral and temporal characteristics of madd duration. This study contributes to the development of computational approaches for Tajweed analysis and supports more objective evaluation of Qur'anic recitation. Future work will focus on dataset expansion, robustness improvement, and real-time implementation.</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

A Two-Stage Hybrid Convolutional Recurrent Neural Network for Hierarchical Detection of Harakat Vowel Lengths in Qur’anic Recitation

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • Nenny Anggraini,
  • Zulkifli Zulkifli,
  • Yusuf Rahman,
  • Achmad Nizar Hidayanto,
  • Husni Teja Sukmana

摘要

The subjective evaluation of vowel elongation (madd) in Qur'anic recitation presents a challenge for standardized assessment. This study proposes a two-stage hybrid convolutional recurrent neural network (CRNN) framework to detect and classify madd occurrences in the recitation of Surah Al-Fatihah. The model is trained on a curated dataset of 480 audio recordings, which are systematically segmented into 8,640 labeled madd instances, comprising 4,800 madd Thabi’i, 3,360 madd ‘Aridh Lissukun, and 480 madd Lazim Mutsaqqal Kilmi. The study investigates eight acoustic features, including mel-frequency cepstral coefficients (MFCCs), formants, and onset energy. An ablation study identifies MFCC as the most critical feature, as its removal reduces classification accuracy to 44.9%, while other features provide complementary information for spectral and vocal tract representation. The proposed hierarchical framework first distinguishes between madd Thabi’i and madd Far’i, followed by fine-grained classification, effectively addressing the limitations observed in single-stage models. Experimental results show that the model achieves 98% accuracy on the test dataset and maintains strong generalization performance on unseen real-world recordings with an overall accuracy of 98.33%. These findings demonstrate that the hybrid CRNN effectively captures both spectral and temporal characteristics of madd duration. This study contributes to the development of computational approaches for Tajweed analysis and supports more objective evaluation of Qur'anic recitation. Future work will focus on dataset expansion, robustness improvement, and real-time implementation.