Elamur Research⌕
☰
home / papers / 2025 / Automatic Pronunciation Error Detection and Correction of the Holy Quran's Learners Using Deep Learning

Automatic Pronunciation Error Detection and Correction of the Holy Quran's Learners Using Deep Learning

Paper · 2025 · Abdullah Abdelfattah, Mahmoud I. Khalil and Hazem Abbas

Summary

The paper tackles automatic detection of recitation mistakes made by learners of the Holy Quran. Existing resources were small, covered only one to three Tajweed rules, or were closed. The authors frame the task as speech recognition into a purpose-built transcription, the Quran Phonetic Script (QPS). QPS is derived from classical scholarship rather than IPA. It has one phoneme level plus ten levels for articulation attributes (Sifat), and it encodes all Hafs Tajweed rules except Ishmam.

To get training data, they built a pipeline that is about 98% automatic:

  • collect recordings from expert reciters,
  • cut them at pause points with a fine-tuned wav2vec2-BERT,
  • transcribe the segments,
  • check each transcript against the known text with their Tasmeea algorithm.

The pipeline yielded 848 hours of labelled audio. A multi-level CTC model trained on this data reaches very low error on expert test data and solid results on a new benchmark of genuine learner mistakes. Code, data and models are open-source. This gives the field a reusable, rule-complete basis for automated tutoring and assessment.

Key points

  • QPS has 11 levels: a 43-symbol phoneme vocabulary and ten Sifat levels with 2–3 values each. A single flat label set for the same information would need 99,072 tokens.
  • The muaalem-annotated-v3 dataset has 286K utterances (848 h) from 22 male expert reciters. The Uthmani-to-phonetic conversion runs as 26 ordered operations.
  • qdat_bench contains 159 real learner recordings (120 female, 39 male) of one verse, Al-Ma'idah 5:109, labelled by a single annotator. It covers phonemes, diacritics, Ghunnah, Qalqalah and Madd.
  • Results for the multi-level CTC model:
    • average PER is 0.21% on the test set and 1.94% on qdat_bench;
    • Tajweed F1 is 75.8% and accuracy on female voices is 84.7%;
    • Ikhfaa is the weakest rule, at F1 0.453.
  • The model has 11 output heads, each with its own CTC loss, combined by a weighted average. Benchmark metrics come with bootstrap confidence intervals (n=10,000).

Interesting ideas

  • I1. Instead of one huge combined label, the model predicts each attribute dimension on a separate head with its own loss, which keeps every head's vocabulary tiny and lets rare combinations be composed.
  • I2. The transcription is generated from expert rules by a fixed chain of small ordered conversion steps, so every output symbol traces back to a named rule.
  • I3. Automatic transcripts are checked against an already known reference text (Tasmeea), so humans only handle the mismatches and the pipeline stays mostly automatic.
  • I4. Elongation length is written as a repeated symbol, so duration becomes a countable quantity and is scored with RMSE instead of right/wrong.
  • I5. The model is trained only on clean expert speech but judged on a small benchmark of genuine learner errors and an unseen speaker group, with bootstrap intervals.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (tokens): the tagger could predict each UD feature (Case, Number, Tense, etc.) on its own head instead of a fused tag, so unseen feature combinations stay reachable and each decision stays explainable.
  • I2 (tree): UD dialect conversion could be organised the same way, as a sequence of small, individually testable rewrite passes with a log of which pass changed what.
  • I3 (learn): silver trees could be compared with a trusted reference or with other witness parsers, and only the disagreements sent to an LLM teacher or a human.
  • I4 (answer): numeric answers in maths word problems could get an error-size metric next to exact match, showing whether failures are near misses or wild guesses.
  • I5 (learn): a compact set of real, hand-checked failure cases plus a held-out text genre with bootstrap confidence intervals would test robustness more honestly than an in-distribution split.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Ain Shams UniversityCairo · Egypt3 works · 8 citations

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 9,086 worksArtificial Intelligence · 8,889 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 worksPsychology · 798 worksExperimental and Cognitive Psychology · 463 worksSignal Processing · 213 works

Cite

Cite as Abdullah Abdelfattah, Mahmoud I. Khalil and Hazem Abbas. 2025. Automatic Pronunciation Error Detection and Correction of the Holy Quran's Learners Using Deep Learning. arXiv:2509.00094.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2509.00094 · arXiv 2509.00094

@misc{abdelfattah2025automatic,
  title = {{Automatic Pronunciation Error Detection and Correction of the Holy Quran's Learners Using Deep Learning}},
  author = {Abdullah Abdelfattah, Mahmoud I. Khalil and Hazem Abbas},
  year = {2025},
  doi = {10.48550/arxiv.2509.00094},
  eprint = {2509.00094},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2509.00094}
}

References

35 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abdou, S. M. and Rashwan, M. A computer aided pronunciation learning system for teaching the holy quran recitation rules. In 2014 IEEE/ACS 11th International Conference on Computer Systems and Applications (AICCSA), pp. 543–550. find
  2. AI, T. Whisper base arabic quran (automatic speech recognition for quranic recitation). https://huggingface.co/tarteel-ai/whisper-base-ar-quran, 2023. Model by Tarteel AI. Company website: https://www.tarteel.ai/. ↗
  3. Al-Dabbaa, A. Clear Expression in the Disputed Words about Hafs. Maktabat wa Matba’at Mustafa al-Babi al-Halabi wa Awladuh. Original: ﺍﺕ ﺍ ﺍ ↗
  4. Al-Marri, M., Raafat, H., Abdallah, M., Abdou, S., and Rashwan, M. Computer aided qur’an pronunciation using dnn. Journal of Intelligent & Fuzzy Systems, 34(5): 3257–3271, 2018. find
  5. Gulati, A., Qin, J., Chiu, C.-C., Parmar, N., Zhang, Y., Yu, J., Han, W., Wang, S., Zhang, Z., Wu, Y., et al. Conformer: Convolution-augmented transformer for speech recognition. arXiv preprint arXiv:2005.08100, 2020. ↗
  6. Al-Swaid, A. R. The Illustrated Tajweed. Dar al-Ghubthani for Quranic Studies. Original: ﺭ ﺍ ﺍ [2021]. find
  7. Hinton, G. E. and Salakhutdinov, R. R. Reducing the dimensionality of data with neural networks. science, 313 (5786):504–507, 2006. find
  8. Alqadasi, A. M. A., Zeki, A. M., Sunar, M. S., Salam, M. S. B. H., Abdulghafor, R., and Khaled, N. A. Improving automatic forced alignment for phoneme segmentation in quranic recitation. IEEE Access, 12:229–244, 2023. find
  9. Hopfield, J. J. Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences, 79(8):2554–2558, 1982. find
  10. Alsahafi, Y. S. and Asad, M. Empirical study on mispronunciation detection for tajweed rules during quran recitation. In 2024 6th International Conference on Computing and Informatics (ICCI), pp. 39–45, 2024. doi: 10.1109/ICCI61671.2024.10485145. ↗
  11. Ibn al Jazari, M. i. M. al-Nashr fı al-Qira’at al-’Ashr. a. Original: ﺍ ﺕﺍﺀﺍ ﺍ ﺍ. Ibn al Jazari, S. a.-D. Al-Jazariyyah Introduction. Dar al-Ghubthani for Quranic Studies, b. Original: ﺭ ﺍ ﺍ. Edited and annotated by Ayman Rushdi Al-Swaid [2021]. find
  12. Baevski, A., Zhou, Y., Mohamed, A., and Auli, M. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449–12460, 2020.
  13. Jordal, I. and Contributors. Audiomentations: A python library for audio data augmentation. https://github.com/iver56/audiomentations, 2025. ↗
  14. Barrault, L., Chung, Y.-A., Meglioli, M. C., Dale, D., Dong, N., Duppenthaler, M., Duquenne, P.-A., Ellis, B., Elsahar, H., Haaheim, J., et al. Seamless: Multilingual expressive and streaming speech translation. arXiv preprint arXiv:2312.05187, 2023. ↗
  15. Khan, H. I., Abid, A., Moussa, M. M., and Abou-Allaban, A. The tarteel dataset: crowd-sourced and labeled quranic recitation. 2021. find
  16. Kheir, Y. E., Ali, A., and Chowdhury, S. A. Automatic pronunciation assessment–a review. arXiv preprint arXiv:2310.13974, 2023. ↗
  17. Chung, Y.-A., Zhang, Y., Han, W., Chiu, C.-C., Qin, J., Pang, R., and Wu, Y. W2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training. In 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pp. 244–250. find
  18. Kheir, Y. E., Ibrahim, O., Meghanani, A., Almarwani, N., Toyin, H. O., Alharbi, S., Alfadly, M., Alkanhal, L., Selim, I., Elbatal, S., et al. Towards a unified benchmark for arabic pronunciation assessment: Quranic recitation as case study. arXiv preprint arXiv:2506.07722, 2025. ↗
  19. Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. Proceedings of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pp. 4171–4186, 2019.
  20. Levenshtein, V. I. Binary codes capable of correcting deletions, insertions, and reversals. Soviet Physics Doklady, 10(8):707–710, 1966. find
  21. Mohammed, A., Sunar, M. S. B., and Salam, M. S. H. Recognition of holy quran recitation rules using phoneme duration. In International Conference of Reliable Information and Communication Technology, pp. 343–352. Springer, 2017. find
  22. Dukes, K. and Habash, N. Morphological annotation of quranic arabic. In Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC 2010), Valletta, Malta, 2010. European Language Resources Association (ELRA).
  23. Omran, D., Fawzi, S., and Kandil, A. Automatic detection of some tajweed rules. In 2023 20th Learning and Technology Conference (L&T), pp. 157–160, 2023a. doi: 10.1109/LT58159.2023.10092350. ↗
  24. Graves, A., Fernández, S., Gomez, F., and Schmidhuber, J. Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks. Omran, D., Fawzi, S., and Kandil, A. Automatic detection of some tajweed rules. In 2023 20th Learning and Technology Conference (L&T), pp. 157–160. find
  25. Osman, H. M., Mustafa, B. S., and Faisal, Y. Qdat: a data set for reciting the quran. International Journal on Islamic Applications in Computer Science And Technology, 9(1):1–9, 2021. find
  26. Park, D. S., Chan, W., Zhang, Y., Chiu, C.-C., Zoph, B., Cubuk, E. D., and Le, Q. V. Specaugment: A simple data augmentation method for automatic speech recognition. arXiv preprint arXiv:1904.08779, 2019. ↗
  27. Plaquet, A. and Bredin, H. Powerset multi-class cross entropy loss for neural speaker diarization. In Proc. INTER-SPEECH 2023, 2023. find
  28. Putra, B., Atmaja, B. T., and Prananto, D. Developing speech recognition system for quranic verse recitation learning software. IJID (International Journal on Informatics for Development), 1(2):1–8, 2012. find
  29. Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., and Sutskever, I. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pp. 28492–28518. PMLR, 2023.
  30. Schneider, S., Baevski, A., Collobert, R., and Auli, M. wav2vec: Unsupervised pre-training for speech recognition. arXiv preprint arXiv:1904.05862, 2019. ↗
  31. Shaiakhmetov, D., Gimaletdinova, G., Momunov, K., and Cankurt, S. Evaluation of the pronunciation of tajweed rules based on dnn as a step towards interactive recitation learning. arXiv preprint arXiv:2503.23470, 2025. ↗
  32. Sherif, M., Samir, A., Khalil, A., and Mohsen, R. Enhancing usability of capl system for quran recitation learning. INTERSPEECH, 2007. find
  33. Team, S. Silero vad: pre-trained enterprise-grade voice activity detector (vad), number detector and language classifier. https://github.com/snakers4/ silero-vad, 2024. ↗
  34. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., and Polosukhin, I. Attention is all you need. Advances in Neural Information Processing Systems, 30:5998–6008, 2017.
  35. ﻱﺭﻭ ﺍ . ﺍ ْ ّ ﺍ ﺙﺍ ِ ّ ﺍ َ ْ َ َ َ ْ َ : ﺭ ﺍ ﺍ ﺍ ﺱ ِ ﺭ ْ ّ ﺍَ ﻭ ِ ِ َ ﺍ. edition, 2008. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/abdelfattah-2025-automatic-pronunciation-error-detection-correction/