Elamur Research⌕
☰
home / papers / 2025 / Saarland-Groningen at NADI 2025 Shared Task: Effective Dialectal Arabic Speech Processing under Data Constraints

Saarland-Groningen at NADI 2025 Shared Task: Effective Dialectal Arabic Speech Processing under Data Constraints

Paper · 2025 · Badr M. Abdullah, Yusser Al Ghussin, Zena Al-Khalili et al.

Summary

The paper describes the Saarland-Groningen team's entries (called BYZÖ) to NADI 2025, the shared task on speech in Arabic dialects. The team entered two subtasks: spoken dialect identification (ADI) and multidialectal ASR. Both cover eight country-level dialects from the Casablanca corpus. For dialect adaptation the team used only the data supplied by the organizers.

  • ADI: They fine-tuned w2v-BERT 2.0 with an 8-way classifier head. They also enlarged the training set with k-NN voice conversion, which re-renders each clip in other speakers' voices.
  • ASR: They built two systems. The first is a CTC model on w2v-BERT 2.0, pre-trained on public Arabic speech corpora and decoded with per-dialect trigram LMs. The second is Whisper-large-v3, tuned first on all dialects together and then once per dialect.

The CTC model clearly beats zero-shot Whisper. After adaptation, however, Whisper comes out ahead. Voice conversion gives a large gain for ADI on a test set from a different domain. Overall the team placed second in ADI and third in ASR. This suggests that careful adaptation can partly make up for having little in-dialect data.

Key points

  • ADI results (test set):
    • The ECAPA-TDNN VoxLingua107 baseline reaches 61.09% accuracy at cost 0.342.
    • Fine-tuned w2v-BERT 2.0 reaches 68.71%, but its cost rises to 1.136.
    • Adding voice conversion lifts accuracy to 76.40% and lowers cost to 0.227.
  • Augmentation setup: Four Arabic-speaking LibriVox voices served as conversion targets, which made the synthetic data four times the size of the original. The ADI splits are 8 hours each; the ASR splits are 12,800 / 12,800 / 10,298 utterances.
  • CTC system (average WER): 58.89 zero-shot, 50.08 after fine-tuning on task data, and 44.40 with Kneser-Ney trigram LMs added. Zero-shot Whisper averages 93.89 and only works on Jordanian (46.09).
  • Whisper system (average WER / CER):
    • Single-stage tuning gives 47.15 WER.
    • Two-stage tuning, stopped on CER, gives the best result: 39.78 WER and 14.76 CER.
    • The best model by WER for a given dialect is not always the best by CER.
  • Persistent gaps: Differences between dialects remain large, with WER from roughly 27.7 to 59.4. Mauritanian and Moroccan are the hardest.

Interesting ideas

  • I1. Changing the voice of each training clip while keeping its dialect label makes the classifier rely less on speaker and recording traits, so it holds up better under domain shift.
  • I2. First adapt one shared model on all varieties pooled together, then split it into variety-specific copies, so every specialist starts from shared knowledge.
  • I3. Using a finer-grained metric (CER) for early stopping than the headline metric (WER) gave the best final scores at little cost.
  • I4. A small, transparent n-gram LM used during decoding removed outputs that sounded plausible but were linguistically unlikely, and it cut WER noticeably.
  • I5. Which architecture is better depended on the regime: the CTC model won zero-shot, but Whisper won after fine-tuning, so judging by only one setting would have misled the team.

For Mova

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

  • I1 (tree): Treebank sentences could be re-rendered with surface variations (spelling, casing, punctuation style) while the gold UD trees stay fixed, which would make the parsers less sensitive to formatting noise.
  • I2 (tree): gparse could be trained on pooled English treebanks and then branched into per-genre or per-dialect variants (web, spoken, academic) from the same checkpoint.
  • I3 (learn): Training could stop on a finer signal, such as per-arc or label accuracy, while LAS is still reported, to avoid overfitting to frequent whole-sentence patterns.
  • I4 (tokens): A light, inspectable n-gram prior could rerank ambiguous tag or lemma candidates without turning the tagger into a black box.
  • I5 (learn): The perceptron parser and gparse should be compared both on unseen genres and after a small adaptation step before deciding which to invest in.

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

🎓Saarland UniversitySaarbrücken · Germany148 works · 18,287 citations🎓University of GroningenGroningen · Netherlands70 works · 5,142 citations🔬German Research Centre for Artificial IntelligenceKaiserslautern · Germany77 works · 29,479 citations🏥University Medical Center GroningenGroningen · Netherlands2 works · 1 citation

 Fields, topics and concepts

📚 Shelves 4The 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 works

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

Cite

Cite as Badr M. Abdullah, Yusser Al Ghussin, Zena Al-Khalili, Ömer Tarik Özyilmaz, Matias Valdenegro-Toro, Simon Ostermann and Dietrich Klakow. 2025. Saarland-Groningen at NADI 2025 Shared Task: Effective Dialectal Arabic Speech Processing under Data Constraints. 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.102

@inproceedings{abdullah2025saarland,
  title = {{Saarland-Groningen at NADI 2025 Shared Task: Effective Dialectal Arabic Speech Processing under Data Constraints}},
  author = {Badr M. Abdullah, Yusser {Al Ghussin}, Zena Al-Khalili, Ömer Tarik Özyilmaz, Matias Valdenegro-Toro, Simon Ostermann and Dietrich Klakow},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-sharedtasks.102},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.102}
}

References

17 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ahmed Amine Ben Abdallah, Ata Kabboudi, Amir Kanoun, and Salah Zaiem. 2024. Leveraging data collection and unsupervised learning for code-switched tunisian arabic automatic speech recognition. In ICASSP 2024-2024 IEEE International Conference On Acoustics, Speech And Signal Processing (ICASSP), pages 12607–12611. IEEE.
  2. Ahmed Abdelali, Hamdy Mubarak, Younes Samih, Sabit Hassan, and Kareem Darwish. 2021. QADI: Arabic dialect identification in the wild. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 1–10, Kyiv, Ukraine (Virtual). Association for Computational Linguistics.
  3. Muhammad Abdul-Mageed, AbdelRahim Elmadany, Chiyu Zhang, El Moatez Billah Nagoudi, Houda Bouamor, and Nizar Habash. 2023. Nadi 2023: The fourth nuanced arabic dialect identification shared task. arXiv preprint arXiv:2310.16117. ↗
  4. Muhammad Abdul-Mageed, Amr Keleg, AbdelRahim Elmadany, Chiyu Zhang, Injy Hamed, Walid Magdy, Houda Bouamor, and Nizar Habash. 2024. Nadi 2024: The fifth nuanced arabic dialect identification shared task. arXiv preprint arXiv:2407.04910. ↗
  5. Muhammad Abdul-Mageed, Chiyu Zhang, Houda Bouamor, and Nizar Habash. 2020. NADI 2020: Muhammad Abdul-Mageed, Chiyu Zhang, Abdel-Rahim A. Elmadany, Houda Bouamor, and Nizar Habash. 2021. NADI 2021: The second nuanced arabic dialect identification shared task. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, WANLP 2021, Kyiv, Ukraine (Virtual), April 9, 2021, pages 244–259. Association for Computational Linguistics.
  6. Badr M Abdullah, Matthew Baas, Bernd Möbius, and Dietrich Klakow. 2025. Voice conversion improves cross-domain robustness for spoken arabic dialect identification. arXiv e-prints, pages arXiv–2505.
  7. Mohammad Al-Fetyani, Muhammad Al-Barham, Gheith Abandah, Adham Alsharkawi, and Maha Dawas. 2023. Masc: Massive arabic speech corpus. In 2022 IEEE Spoken Language Technology Workshop (SLT), pages 1006–1013. find
  8. Karim Al-Haff, Mustafa Jarrar, Tymaa Hammouda, and Fadi Zaraket. 2022. Curras + baladi: Towards a Levantine corpus. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 769–778, Marseille, France. European Language Resources Association. ↗
  9. Sadeen Alharbi, Areeb Alowisheq, Zoltán Tüske, Kareem Darwish, Abdullah Alrajeh, Abdulmajeed Alrowithi, Aljawharah Bin Tamran, Asma Ibrahim, Raghad Aloraini, Raneem Alnajim, and 1 others. 2024. Sada: Saudi audio dataset for arabic. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 10286–10290. IEEE. find
  10. Anass Allak, Naira Abdou Mohamed, Imade Benelallam, and Kamel Gaanoun. 2021. Dialectal voice : An open-source voice dataset and automatic speech recognition model for moroccan arabic dialect. In Proceedings of the Data Centric AI (NeurIPS 2021). find
  11. R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber. 2020. Common voice: A massively-multilingual speech corpus. In Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4211–4215. ↗
  12. Matthew Baas, Benjamin van Niekerk, and Herman Kamper. 2023. Voice conversion with just nearest neighbors. In Interspeech 2023, pages 2053–2057. find
  13. Omar A Essameldin, Ali O Elbeih, Wael H Gomaa, and Wael F Elsersy. 2025. Arabic dialect classification using rnns, transformers, and large language models: A comparative analysis. arXiv preprint arXiv:2506.19753. ↗
  14. Nizar Y Habash. 2010. Introduction to Arabic natural language processing. Morgan & Claypool Publishers. ↗
  15. Peter Sullivan, AbdelRahim Elmadany, and Muhammad Abdul-Mageed. 2023. On the robustness of arabic speech dialect identification. In Interspeech 2023, pages 5326–5330.
  16. 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. ↗
  17. 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. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/abdullah-2025-saarland-groningen-nadi-2025-shared/