Elamur Research⌕
☰
home / papers / 2025 / Munsit at NADI 2025 Shared Task 2: Pushing the Boundaries of Multidialectal Arabic ASR with Weakly Supervised Pretraining and Continual Supervised Fine-tuning

Munsit at NADI 2025 Shared Task 2: Pushing the Boundaries of Multidialectal Arabic ASR with Weakly Supervised Pretraining and Continual Supervised Fine-tuning

Paper · 2025 · Mahmoud Salhab, Shameed Sait, Mohammad Abusheikh et al.

Summary

The paper describes the winning system for NADI 2025 Shared Task 2, which evaluated speech recognition across many Arabic dialects. The core difficulty is that human-transcribed Arabic speech is scarce, especially for regional varieties, while dialects differ a great deal from Modern Standard Arabic (MSA). The authors train a large Conformer model with a CTC objective in two phases.

First, the model learns from about 15,000 hours of automatically labelled MSA and dialectal audio, using a labelling pipeline from their earlier work. Second, they continue training at a lower learning rate on a cleaner mix. This mix combines a 3,000-hour quality-filtered slice of the weak data, with news (mostly MSA) removed, and the official Casablanca training set enlarged through noise and silence augmentation.

The system ranked first on both metrics, with an average WER of 35.69% and CER of 12.21% on the blind test. This suggests that curated noisy labels plus a short supervised phase can partly replace expensive manual transcription for dialect-rich, low-resource languages.

Key points

  • Model: Conformer (large variant) trained with CTC, with 18 encoder layers, hidden size 512, 8 heads, conv kernel 31, and about 121M parameters. It uses a SentencePiece vocabulary of only 128 tokens and 80-dim mel features.
  • Training ran on 8 A100 GPUs with a global batch of 512, AdamW, a Noam schedule (10k warm-up, peak LR 2e-3), and bfloat16. Fine-tuning used one tenth of that learning rate.
  • Fine-tuning data was a 3,000-hour filtered weak subset that excluded news/MSA, plus the Casablanca training set augmented with Gaussian noise, background noise and inserted silence.
  • Test results by dialect:
    • Best: Jordanian 20.68% WER, Egyptian 20.89%, Emirati 22.67%.
    • Hardest: Mauritanian 59.03% and Algerian 53.62%.
    • The official baseline averaged 93.90% WER.
  • Average WER moved from 36.83% in the evaluation phase to 35.69% on the blind test. CER rose slightly, from 11.94% to 12.21%.

Interesting ideas

  • I1. Pretrain on a large pool of automatically labelled data, then continue from those weights on a small trusted set with a tenfold lower learning rate, rather than training once on everything mixed together.
  • I2. During fine-tuning, the over-represented "easy" domain (news in MSA) was deliberately removed from the weak data, which pushes the model toward the under-served target varieties.
  • I3. The small gold set was not used alone in the second phase; a quality-filtered slice of the weak data was kept alongside it, which likely limits overfitting and forgetting.
  • I4. The scarce gold data was stretched with cheap perturbations (added noise, inserted silence) that keep the labels valid.
  • I5. Results were broken down per dialect and compared between the development phase and a blind test phase, which reveals both weak varieties and whether gains generalise.

For Mova

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

  • I1 (learn): gparse could first train on silver trees agreed by witness parsers, then take a short low-LR pass on gold UD treebanks.
  • I2 (learn): when adapting toward fables or school texts, newswire-heavy silver could be down-filtered so the parser and tagger spend capacity on the target genres.
  • I3 (tree): a fine-tuning batch for a new UD dialect conversion could blend converted gold trees with high-agreement silver trees from the original corpus.
  • I4 (tokens): the tokenizer and tagger could be trained on label-preserving text perturbations such as casing changes, spacing and punctuation noise, or common typos, to harden them against messy input.
  • I5 (answer): QA and test scores could be reported per genre (fables, maths problems, school courses), with a held-out blind split, so regressions on one text type are not hidden by an average.

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

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSignal Processing · 452 works

🧭 Topics 2Research subjects and languages the work is about.Semitic languages · 2,908 worksSpeech & spoken language · 480 works

Cite

Cite as Mahmoud Salhab, Shameed Sait, Mohammad Abusheikh and Hasan Abusheikh. 2025. Munsit at NADI 2025 Shared Task 2: Pushing the Boundaries of Multidialectal Arabic ASR with Weakly Supervised Pretraining and Continual Supervised Fine-tuning. Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.arabicnlp-sharedtasks.100

@inproceedings{salhab2025munsit,
  title = {{Munsit at NADI 2025 Shared Task 2: Pushing the Boundaries of Multidialectal Arabic ASR with Weakly Supervised Pretraining and Continual Supervised Fine-tuning}},
  author = {Mahmoud Salhab, Shameed Sait, Mohammad Abusheikh and Hasan Abusheikh},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-sharedtasks.100},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.100}
}

References

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

  1. Mahmoud Al-Ayyoub, Aya Nuseir, Kholoud Alsmearat, Yaser Jararweh, and Brij Gupta. 2018. Deep learning for arabic nlp: A survey. Journal of Computational Science, 26:522–531. ↗
  2. Mohammad Al-Fetyani, Muhammad Al-Barham, Gheith Abandah, Adham Alsharkawi, and Maha Dawas. 2021. Masc: Massive arabic speech corpus. find
  3. Wajdan Algihab, Noura Alawwad, Anfal Aldawish, and Sarah AlHumoud. 2019. Arabic speech recognition with deep learning: A review. In Social Computing and Social Media. Design, Human Behavior and Analytics, pages 15–31, Cham. Springer International Publishing. find
  4. Sadeen Alharbi, Areeb Alowisheq, Zoltán Tüske, Kareem Darwish, Abdullah Alrajeh, Abdulmajeed Alrowithi, Aljawharah Bin Tamran, Asma Ibrahim, Raghad Aloraini, Raneem Alnajim, Ranya Alkahtani, Renad Almuasaad, Sara Alrasheed, Shaykhah Alsubaie, and Yaser Alonaizan. 2024. Sada: Saudi audio dataset for arabic. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 10286–10290. find
  5. Ahmed Ali, Hamdy Mubarak, and Stephan Vogel. 2014. Advances in dialectal arabic speech recognition: A study using twitter to improve egyptian asr. In Proceedings of the 11th International Workshop on Spoken Language Translation: Papers, pages 156–162.
  6. Ahmed Alwajeeh, Mahmoud Al-Ayyoub, and Ismail Hmeidi. 2014. On authorship authentication of arabic articles. In 2014 5th International Conference on Information and Communication Systems (ICICS), pages 1–6. find
  7. L Bouchakour and M Debyeche. 2018. Improving continuous arabic speech recognition over mobile networks dsr and nsr using mfccs features transformed. International Journal of Circuits, Systems and Signal Processing, 12:1–8. find
  8. Paul Cardinal, Ahmed Ali, Najim Dehak, Yifan Zhang, Takahiro A. Hanai, Yu Zhang, James R. Glass, and Stephan Vogel. 2014. Recent advances in asr applied to an arabic transcription system for al-jazeera. In Proceedings of Interspeech 2014, pages 2088–2092. find
  9. Frank Diehl, Mark JF Gales, Marcus Tomalin, and Philip C Woodland. 2012. Morphological decomposition in arabic asr systems. Computer Speech & Language, 26(4):229–243. ↗
  10. Amirbek Djanibekov, Hawau Olamide Toyin, Raghad Alshalan, Abdullah Alatir, and Hanan Aldarmaki. 2025. Dialectal coverage and generalization in Arabic speech recognition. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 29490– 29502, Vienna, Austria. Association for Computational Linguistics. ↗
  11. Dongji Gao, Hainan Xu, Desh Raj, Leibny Paola Garcia Perera, Daniel Povey, and Sanjeev Khudanpur. 2023. Learning from flawed data: Weakly supervised automatic speech recognition. In 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 1–8. IEEE. find
  12. Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. 2006. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369–376. find
  13. Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang. 2020. Conformer: Convolution-augmented transformer for speech recognition. Preprint, arXiv:2005.08100. ↗
  14. Hamza Kheddar, Mustapha Hemis, and Yassine Himeur. 2024. Automatic speech recognition using advanced deep learning approaches: A survey. Information Fusion, page 102422. ↗
  15. Taku Kudo and John Richardson. 2018. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. Preprint, arXiv:1808.06226. ↗
  16. Liang Lu, Changliang Liu, Jinyu Li, and Yifan Gong. 2020. Exploring transformers for large-scale speech recognition. arXiv preprint arXiv:2005.09684. ↗
  17. Mohamed Amine Menacer, Odile Mella, Dominique Fohr, Denis Jouvet, David Langlois, and Kamel Smaïli. 2017. Development of the arabic loria automatic speech recognition system (alasr) and its evaluation for algerian dialect. Procedia Computer Science, 117:81–88. Arabic Computational Linguistics. find
  18. Amr Mousa, Hong-Kwang Kuo, Lidia Mangu, and Hagen Soltau. 2013. Morpheme-based feature-rich language models using deep neural networks for lvcsr of egyptian arabic. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8435–8439. find
  19. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023a. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492–28518. PMLR. ↗
  20. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine Mcleavey, and Ilya Sutskever. 2023b. Robust speech recognition via large-scale weak supervision. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 28492–28518. PMLR. ↗
  21. Mahmoud Salhab, Marwan Elghitany, Shameed Sait, Syed Sibghat Ullah, Mohammad Abusheikh, and Hasan Abusheikh. 2025. Advancing arabic speech recognition through large-scale weakly supervised learning. Preprint, arXiv:2504.12254. Bashar Talafha, Karima Kadaoui, Samar Mohamed Magdy, Mariem Habiboullah, Chafei Mohamed Chafei, Ahmed Oumar El-Shangiti, Hiba Zayed, Mohamedou cheikh tourad, Rahaf Alhamouri, Rwaa Assi, Aisha Alraeesi, Hour Mohamed, Fakhraddin Alwajih, Abdelrahman Mohamed, Abdellah El Mekki, El Moatez Billah Nagoudi, Benelhadj Djelloul Mama Saadia, Hamzah A. Alsayadi, Walid Al-Dhabyani, and 8 others. 2024. Casablanca: Data and models for multidialectal arabic speech recognition. Preprint, arXiv:2410.04527. ↗
  22. Bashar Talafha, Hawau Olamide Toyin, Peter Sullivan, AbdelRahim Elmadany, Abdurrahman Juma, Amirbek Djanibekov, Chiyu Zhang, Hamad Alshehhi, Hanan Aldarmaki, Mustafa Jarar, Nizar Habash, and Muhammad Abdul-Mageed. 2025. Nadi 2025: The first multidialectal arabic speech processing shared task. In The Third Arabic Natural Language Processing Conference (ArabicNLP 2025), Suzhou. Association for Computational Linguistics. ↗
  23. Jayashri Vajpai and Avnish Bora. 2016. Industrial applications of automatic speech recognition systems. International Journal of Engineering Research and Applications, 6(3):88–95. find
  24. Abdul Waheed, Bashar Talafha, Peter Sullivan, Abdel-Rahim Elmadany, and Muhammad Abdul-Mageed. 2023. Voxarabica: A robust dialect-aware arabic speech recognition system. arXiv preprint arXiv:2310.11069. ↗
  25. Yongqiang Wang, Yangyang Shi, Frank Zhang, Chunyang Wu, Julian Chan, Ching-Feng Yeh, and Alex Xiao. 2021. Transformer in action: A comparative study of transformer-based acoustic models for large scale speech recognition applications. In ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6778–6782. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/salhab-2025-munsit-nadi-2025-shared-task/