Elamur Research⌕
☰
home / papers / 2025 / ELYADATA & LIA at NADI 2025: ASR and ADI Subtasks

ELYADATA & LIA at NADI 2025: ASR and ADI Subtasks

Paper · 2025 · Haroun Elleuch, Youssef Saidi, Salima Mdhaffar et al.

Summary

The Elyadata–LIA team took part in two speech subtasks of the NADI 2025 shared task: spoken Arabic Dialect Identification (ADI) and multi-dialectal ASR. Both cover eight country-level dialects. Arabic dialects vary widely and lack a fixed spelling, so systems must cope with large variation.

For ADI, the team built a classifier from the Whisper-large-v3 encoder, an attention pooling layer and a softmax output. The lower 16 encoder layers were frozen, and training used several data augmentations. Training happened in two stages: first on the balanced ADI-20-53h corpus, then on the small NADI adaptation split. This system ranked first, with 79.83% test accuracy.

For ASR, the team compared Whisper-medium and Whisper-large-v3, and joint versus per-dialect training. They then chose SeamlessM4T-v2 Large (Egyptian variant), fine-tuned once for each dialect. It ranked second, with an average test WER of 38.54% and CER of 14.53%.

The work shows that large pretrained speech models adapt well to dialects through careful data choices and specialization.

Key points

  • First-stage data matters for zero-shot ADI. On the NADI validation set, a model trained on ADI-17 reached only 31.84%. Training on ADI-20-53h raised this to 78.33%. ADI-20-53h holds at most 53 h per variety (1,060 h in total) across 20 varieties, including MSA.
  • Final ADI results: 98.08% validation and 79.83% test accuracy, with NIST LRE 2022 average costs of 0.0171 and 0.1788. Algerian was the hardest dialect and was mostly confused with Moroccan. Egyptian and Jordanian were also confused with each other, which the authors found surprising.
  • NADI ASR data: 1,600 utterances per dialect in each split, about 15.4 h for training and 15.3 h for validation.
  • Whisper results: On the combined data, Whisper-medium clearly beat Whisper-large-v3 (48.21% vs 72.20% WER). Per-dialect Whisper-medium models lowered WER to 46.71%, while CER stayed roughly the same.
  • SeamlessM4T-v2 Large per-dialect models beat Whisper on every dialect in validation (39.17% average WER). On the test set, Levantine and Egyptian dialects were easiest and Maghrebi dialects hardest; Mauritanian had the highest WER at 58.11%.

Interesting ideas

  • I1. Adapting in two steps, first on a broad related corpus and then on a tiny in-domain split, worked far better than either step alone, and the choice of first-stage corpus decided most of the zero-shot quality.
  • I2. A capped, per-class-balanced subset (at most 53 h per variety) was preferred over the full, skewed corpus, since even coverage beat raw volume.
  • I3. Freezing the lower encoder layers keeps general representations intact, while only the upper layers learn the new task.
  • I4. Bigger was not better: the medium Whisper model clearly beat the large one after fine-tuning, so model size has to be tested empirically rather than assumed.
  • I5. Reading the confusion matrix revealed both expected confusions (neighbouring Maghrebi dialects) and unexpected ones (Egyptian with Jordanian), which point to data or model artefacts.

For Mova

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

  • I1 (learn): Pretrain parsers on broad silver data agreed by witnesses, then adapt briefly on small gold sets such as fables, checking zero-shot scores before adaptation to judge the base corpus.
  • I2 (tree): When mixing UD treebanks or genres for gparse, cap each source's share so one large treebank does not dominate the annotation dialect.
  • I3 (tree): When fine-tuning the transformer in gparse on a new domain, freezing the lower layers could limit forgetting and reduce how much labelled data is needed.
  • I4 (learn): Always evaluate smaller model or teacher tiers on the same setup, since a cheaper option may win and simplify the pipeline.
  • I5 (frames): Build confusion matrices over semantic roles or speech-act labels to find systematic mix-ups that should become targeted repair rules.

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

🔬Laboratoire Informatique d'AvignonAvignon · France7 works · 436 citations🎓Université d'Avignon et des Pays de VaucluseAvignon · France6 works · 603 citations

 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 Haroun Elleuch, Youssef Saidi, Salima Mdhaffar, Yannick Estève and Fethi Bougares. 2025. ELYADATA & LIA at NADI 2025: ASR and ADI Subtasks. 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.105

@inproceedings{elleuch2025elyadata,
  title = {{ELYADATA \& LIA at NADI 2025: ASR and ADI Subtasks}},
  author = {Haroun Elleuch, Youssef Saidi, Salima Mdhaffar, Yannick Estève and Fethi Bougares},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-sharedtasks.105},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.105}
}

References

15 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Bashar Alhafni, Sarah Al-Towaity, Ziyad Fawzy, Fatema Nassar, Fadhl Eryani, Houda Bouamor, and Nizar Habash. 2024. Exploiting dialect identification in automatic dialectal text normalization. In Proceedings of the Second Arabic Natural Language Processing Conference, pages 42–54, Bangkok, Thailand. Association for Computational Linguistics.
  2. Ahmed Ali, Stephan Vogel, and Steve Renals. 2017. Speech recognition challenge in the wild: Arabic mgb-3. In 2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 316– 322. IEEE.
  3. Loïc Barrault, Yu-An Chung, Mariano Coria Meglioli, David Dale, Ning Dong, Mark Duppenthaler, Paul-Ambroise Duquenne, Brian Ellis, Hady Elsahar, Justin Haaheim, and 1 others. 2023. Seamless: Multilingual expressive and streaming speech translation. arXiv preprint arXiv:2312.05187. ↗
  4. Houda Bouamor, Nizar Habash, and Kemal Oflazer. 2014. A multidialectal parallel corpus of Arabic. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC’14), pages 1240–1245, Reykjavik, Iceland. European Language Resources Association (ELRA).
  5. Amira Dhouib, Achraf Othman, Oussama El Ghoul, Mohamed Koutheair Khribi, and Aisha Al Sinani. 2022. Arabic automatic speech recognition: A systematic literature review. Applied Sciences, 12(17). ↗
  6. Haroun Elleuch, Salima Mdhaffar, Yannick Estève, and Fethi Bougares. 2025. Adi-20: Arabic dialect identification dataset and models. In Proceedings of the Annual Conference of the International Speech Communication Association (Interspeech), Rotterdam Ahoy Convention Centre, Rotterdam, The Netherlands. To appear.
  7. Charles A. Ferguson. 1959. 15(2):325–340. Diglossia. WORD, pages 711–718, Istanbul, Turkey. European Language Resources Association (ELRA). find
  8. Nizar Habash, Fadhl Eryani, Salam Khalifa, Owen Rambow, Dana Abdulrahim, Alexander Erdmann, Reem Faraj, Wajdi Zaghouani, Houda Bouamor, Nasser Zalmout, Sara Hassan, Faisal Al-Shargi, Sakhar Alkhereyf, Basma Abdulkareem, Ramy Eskander, Mohammad Salameh, and Hind Saddiki. 2018. Unified guidelines and resources for Arabic dialect orthography. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).
  9. James Lane. 2025. The 10 most spoken languages in the world in 2025. https://www.babbel.com/en/magazine/ the-10-most-spoken-languages-in-the-world. Accessed: 2025-08-12. ↗
  10. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492–28518. PMLR. ↗
  11. Mirco Ravanelli, Titouan Parcollet, Adel Moumen, Sylvain de Langen, Cem Subakan, Peter Plantinga, Yingzhi Wang, Pooneh Mousavi, Luca Della Libera, Artem Ploujnikov, Francesco Paissan, Davide Borra, Salah Zaiem, Zeyu Zhao, Shucong Zhang, Georgios Karakasidis, Sung-Lin Yeh, Pierre Champion, Aku Rouhe, and 14 others. 2024. Open-source conversational AI with SpeechBrain 1.0. Journal of Machine Learning Research. find
  12. Suwon Shon, Ahmed Ali, Younes Samih, Hamdy Mubarak, and James Glass. 2020. Adi17: A fine-grained arabic dialect identification dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8244–8248. IEEE. find
  13. Bashar Talafha, Karima Kadaoui, Samar Mohamed Magdy, Mariem Habiboullah, Chafei Mohamed Chafei, Ahmed Oumar El-Shangiti, Hiba Zayed, Rahaf Alhamouri, Rwaa Assi, Aisha Alraeesi, and 1 others. 2024. Casablanca: Data and models for multidialectal arabic speech recognition. arXiv preprint arXiv:2410.04527. ↗
  14. 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. ↗
  15. Nizar Habash, Mona Diab, and Owen Rambow. 2012. Conventional orthography for dialectal Arabic. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12),

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/elleuch-2025-elyadata-lia-nadi-2025-asr/