Elamur Research⌕
☰
home / papers / 2025 / ASR Models for Traditional Emirati Arabic: Challenges, Adaptations, and Performance Evaluation

ASR Models for Traditional Emirati Arabic: Challenges, Adaptations, and Performance Evaluation

Paper · 2025 · Maha Alblooki, Kentaro Inui and Shady Shehata

Summary

Traditional Emirati Arabic carries Bedouin, coastal and mountain features and heritage vocabulary, but current ASR systems handle it poorly. The authors built a small speech corpus from the Alsanaa heritage radio program and its companion book by Abdullah Bin Dalmook. They aligned audio and text by hand, normalized the transcripts and split the data for training and testing. They then compared five pre-trained models (Wav2Vec 2.0, XLS-R, Whisper Small, Whisper Medium, MMS) without adaptation and after partial fine-tuning, which updated only the top layers. MMS performed best after adaptation, at about 41% WER and 13% CER. Wav2Vec 2.0 came close behind. The XLS-R and Whisper variants stayed far worse, and Whisper Small exceeded 100% WER because of looping output. Error analysis connects the failures to dialect phonology, morphology and lexicon, and to weak Gulf Arabic coverage in pre-training. The work is a first public benchmark and dataset for this dialect, with value for preserving oral heritage.

Key points

  • The corpus has 102 MP3 files (about 4 hours) from one male speaker. The audio comes from Alsanaa, broadcast on Aloula with support from the Hamdan bin Mohammed Heritage Center. It is split 80/10/10, and the data and preprocessing code are released on GitHub (alsanaa-emirati-dataset).

  • Preprocessing removed diacritics and punctuation, cropped non-speech intros and outros, converted audio to 16 kHz mono and normalized amplitude.

  • The checkpoints were:

    • eabayed Wav2Vec2 Emirati model (315M parameters)
    • jonatasgrosman XLS-R-53 Arabic (315M)
    • Whisper Small Arabic (241M)
    • Whisper Medium Arabic (763M)
    • facebook/mms-1b-all (965M)

  • Fine-tuning used AdamW with warm-up then decay, early stopping, speed perturbation, SpecAugment and gradient accumulation.

  • Fine-tuned test WER/CER:

    ModelWER (%)CER (%)
    MMS41.0413.34
    Wav2Vec 2.044.315.96
    Whisper Medium88.6072.03
    XLS-R89.7842.31
    Whisper Small100.0475.53

Interesting ideas

  • I1. The paper catalogues dialect traits by layer (phonological, morphological, lexical, syntactic, fillers, discourse markers), giving one transcript example for each, so errors can be traced to a specific linguistic level.
  • I2. Reporting character-level and word-level error side by side shows where the problem lies: a large gap means sounds are captured but words are segmented or rebuilt wrongly.
  • I3. The authors compare unadapted and adapted performance per model. A strong starting score predicted small gains, while a middling one predicted large gains, so adaptability is treated as a selection criterion in its own right.
  • I4. A WER above 100% was traced to decoder repetition loops and truncation of long inputs, a failure mode that inflates metrics and erases content.
  • I5. Residual errors were inspected for systematic substitution of dialect forms with those of better-represented varieties, exposing pre-training bias.

For Mova

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

  • I1 (tokens): A similar layered inventory of English variety and register quirks would help. It could cover contractions, dialect spellings and fillers, each with corpus examples, to show which tagger or lemma errors belong to which layer.
  • I2 (tokens): Scoring the tokenizer and lemmatizer at both character and token granularity would separate boundary mistakes from genuine misidentification of a form.
  • I3 (learn): When choosing witness parsers or teacher tiers, record both the out-of-the-box score and the gain after brief adaptation, rather than ranking on one number.
  • I4 (back): Generation and the text → graph → text compressor could include an explicit check for repeated n-gram loops and for dropped content relative to the meaning graph.
  • I5 (tree): When converting or parsing non-standard treebanks, check whether errors consistently map dialect constructions onto standard-English analyses, and treat that pattern as a signal of data bias.

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 Maha Alblooki, Kentaro Inui and Shady Shehata. 2025. ASR Models for Traditional Emirati Arabic: Challenges, Adaptations, and Performance Evaluation. Proceedings of the 8th International Conference on Natural Language and Speech Processing (ICNLSP-2025).

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{alblooki2025asr,
  title = {{ASR Models for Traditional Emirati Arabic: Challenges, Adaptations, and Performance Evaluation}},
  author = {Maha Alblooki, Kentaro Inui and Shady Shehata},
  booktitle = {Proceedings of the 8th International Conference on Natural Language and Speech Processing (ICNLSP-2025)},
  year = {2025},
  doi = {10.82419/263},
  url = {https://doi.org/10.82419/263}
}

References

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

  1. 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
  2. Maryam Al Ali and Hanan Aldarmaki. 2024. Mixat: A data set of bilingual emirati-english speech. ↗
  3. Hamzah A. Alsayadi, Salah Al-Hagree, Fahd A. Alqasemi, and Abdelaziz A. Abdelhamid. 2022. Dialectal arabic speech recognition using cnn-lstm based on end-to-end deep learning. In 2022 2nd International Conference on Emerging Smart Technologies and Applications (eSmarTA), pages 1–8. find
  4. Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, and Michael Auli. 2021. XLS-R: self-supervised cross-lingual speech representation learning at scale. CoRR, abs/2111.09296. ↗
  5. Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. CoRR, abs/2006.11477. ↗
  6. Martijn Bartelds, Nay San, Bradley McDonnell, Dan Jurafsky, and Martijn Wieling. 2023. Making more of little data: Improving low-resource automatic speech recognition using data augmentation. find
  7. Abdullah Hamdan Bin Dalmook. 2021. Alsanaa. Hamdan bin Mohammad Heritage Center, Dubai. Paper Cover; Dimensions: 24x17 cm. find
  8. 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). ↗
  9. Mohamed Elmahdy, Mark Hasegawa-Johnson, and Eiman Mustafawi. 2014. Development of a TV broadcasts speech recognition system for qatari Arabic. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC’14), pages 3057–3061, Reykjavik, Iceland. European Language Resources Association (ELRA).
  10. Injy Hamed, Pavel Denisov, Chia-Yu Li, Mohamed Elmahdy, Slim Abdennadher, and Ngoc Thang Vu. 2022. Investigations on speech recognition systems for low-resource dialectal arabic–english code-switching speech. Computer Speech Language, 72:101278. ↗
  11. Mohanad Khudhair and Ahmed Talib. 2022. Improving low resources arabic speech recognition using data augmentation. In 2022 Fifth College of Science International Conference of Recent Trends in Information Technology (CSCTIT), pages 60–65. find
  12. Hamdy Mubarak, Amir Hussein, Shammur Absar Chowdhury, and Ahmed Ali. 2021. Qasr: Qcri aljazeera speech resource – a large scale annotated arabic speech corpus. ↗
  13. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust speech recognition via large-scale weak supervision. ↗
  14. Sakriani Sakti and Benita Angela Titalim. 2023. Leveraging the multilingual indonesian ethnic languages dataset in self-supervised models for low-resource asr task. In 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 1–8. find
  15. 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, Sara Shatnawi, Yasir Ech-Chammakhy, Amal Makouar, Yousra Berrachedi, Mustafa Jarrar, Shady Shehata, Ismail Berrada, and Muhammad Abdul-Mageed. 2024. Casablanca: Data and models for multidialectal arabic speech recognition.
  16. Bashar Talafha, Abdul Waheed, and Muhammad Abdul-Mageed. 2023. N-shot benchmarking of whisper on diverse arabic speech recognition.
  17. Abdul Waheed, Bashar Talafha, Peter Sullivan, Abdel-Rahim Elmadany, and Muhammad Abdul-Mageed. 2023. Voxarabica: A robust dialect-aware arabic speech recognition system.
  18. Ara Yeroyan and Nikolay Karpov. 2024. Enabling asr for low-resource languages: A comprehensive dataset creation approach. ↗
  19. Hiba Adreese Younis and Yusra Faisal Mohammad. 2023. Arabic speech recognition based on self supervised learning. In 2023 16th International Conference on Developments in eSystems Engineering (DeSE), pages 528–533. find
  20. Yunbo Zhang, Deepak Gopinath, Yuting Ye, Jessica Hodgins, Greg Turk, Jungdam Won, Harrison Jesse Smith, Qingyuan Zheng, Yifei Li, Somya Jain, Jessica K. Hodgins, Simran Arora, Patrick Lewis, Angela Fan, Jacob Kahn, Christopher Ré, and Michael Auli. 2023. Scaling speech technology to 1,000+ languages. arXiv preprint arXiv:2305.13516. ↗
  21. Jing Zhao and Wei-Qiang Zhang. 2022. Improving automatic speech recognition performance for low-resource languages with self-supervised models. IEEE Journal of Selected Topics in Signal Processing, 16(6):1227–1241. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/alblooki-2025-asr-models-traditional-emirati-arabic/