Elamur Research⌕
☰
home / papers / 2026 / Parameter-Efficient Adaptation of Self-Supervised Models for Arabic Speech Recognition

Parameter-Efficient Adaptation of Self-Supervised Models for Arabic Speech Recognition

Paper · 2026 · Wafa Mohammed Alshehri, Wasfi G. Al-khatib and Mohammad Ismail Amro

Summary

The paper asks whether large self-supervised speech encoders can be adapted to Arabic speech recognition without updating all their weights. The authors start from wav2vec2-XLSR-53, a model of about 317M parameters, and train it with CTC loss on Mozilla Common Voice Arabic V24.0. This corpus has roughly 92 validated hours of crowdsourced, noisy, accent-diverse recordings, and this is the first study to report results on that release. They compare three regimes under one shared hyperparameter setup:

  • full fine-tuning,
  • LoRA,
  • DoRA.

Transcripts are normalized by removing diacritics and punctuation and by merging alef variants. They are then modelled with a 39-symbol character vocabulary. Full fine-tuning gives the best XLS-R Arabic result the authors know of, at 23.03% WER, without any extra corpora. LoRA reaches 36.10% WER while training about 2% of the parameters. DoRA trails at 45.20% WER. The study is the first to bring LoRA to a CTC-based SSL model for Arabic, and the first to test DoRA for Arabic speech on any architecture. It offers practical guidance for teams with limited compute, and the models and code are released.

Key points

  • Data and splits: Common Voice Arabic V24.0 with its official splits (28,881 train / 10,181 dev / 10,508 test). Audio is resampled to 16 kHz mono, and the CNN feature extractor stays frozen in every regime.
  • Full fine-tuning results: 23.03% WER and 6.7 CER. This beats elgeish/xlsr-53-arabic (26.55%) by 3.52 points, although that model used supplementary data. The earlier figures of 31.16% WER (Talafha et al., CV 11.0) and 40% (Younis et al.) are also surpassed.
  • LoRA: 7.1M trainable parameters (2.2%), a 27 MB adapter, 47× less storage, 36.10% WER and 9.6 CER.
  • DoRA: 7.3M parameters (2.3%), a 27.9 MB adapter, 45.20% WER and 12.54 CER. It still beats some fully fine-tuned community checkpoints, for example kmfoda/xlsr-arabic at 52.53%.
  • Shared configuration:
    • LoRA/DoRA use rank 16, alpha 32 and dropout 0.05.
    • Adapters sit on q_proj, k_proj, v_proj, out_proj, intermediate_dense and output_dense.
    • Training uses a learning rate of 3e-4, an effective batch of about 64 and early stopping on validation WER with patience 3.
    • Training ran for 50 epochs (full FT), 41 (LoRA) and 33 (DoRA).

Interesting ideas

  • I1. Holding every hyperparameter fixed isolates the effect of the adaptation method, but the authors admit this can unfairly penalize a method that needs its own tuning, as DoRA likely did. A rival stopping early (33 vs 41 epochs) is read as a sign of premature stopping, not as an efficiency win.
  • I2. One frozen base model can serve many specialized tasks through small swappable low-rank adapters (about 27 MB each), instead of storing a full copy per task.
  • I3. Before modelling, the authors shrink the symbol inventory by stripping optional marks and collapsing orthographic variants of one letter into a single form. This lets the model focus on the core mapping rather than surface noise.
  • I4. The results table groups baselines by what data they were trained on (extra corpora vs Common Voice only), so a gain cannot be mistaken for a data advantage.

For Mova

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

  • I1 (learn): When comparing training variants, keep one shared configuration as the controlled baseline. Also log the number of epochs to early stop, so that a fast stop gets investigated before any variant is declared worse.
  • I2 (tree): gparse could keep one shared transformer encoder and attach light adapters per genre or treebank dialect (e.g., fables vs school texts), so specialized parsers cost little storage.
  • I3 (tokens): The en tokenizer could map typographic variants (quote styles, dash types, ligatures) to canonical forms early, while keeping the original span so every step stays explainable.
  • I4 (answer): Score reports on school tests or fable QA could list, next to each competing system, which training resources it used, so comparisons stay honest.

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

🎓King Fahd University of Petroleum and MineralsDhahran · Saudi Arabia52 works · 4,666 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 Wafa Mohammed Alshehri, Wasfi G. Al-khatib and Mohammad Ismail Amro. 2026. Parameter-Efficient Adaptation of Self-Supervised Models for Arabic Speech Recognition. Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.abjadnlp-1.40

@inproceedings{alshehri2026parameter,
  title = {{Parameter-Efficient Adaptation of Self-Supervised Models for Arabic Speech Recognition}},
  author = {Wafa Mohammed Alshehri, Wasfi G. Al-khatib and Mohammad Ismail Amro},
  booktitle = {Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script},
  year = {2026},
  doi = {10.18653/v1/2026.abjadnlp-1.40},
  url = {https://doi.org/10.18653/v1/2026.abjadnlp-1.40}
}

References

17 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. This paper presented the first application of Sadeen Alharbi, Areeb Alowisheq, Zoltán Tüske, PEFT methods to CTC-based self-supervised Kareem Darwish, Abdullah Alrajeh, Abdulmajeed Alrowithi, Aljawharah Bin Tamran, Asma models for Arabic ASR using XLS-R. We eval-Ibrahim, Raghad Aloraini, Raneem Alnajim, uated full fine-tuning, LoRA, and DoRA on Ranya Alkahtani, Renad Almuasaad, Sara Al-Common Voice Arabic v24.0. Full fine-tuning rasheed, Shaykhah Alsubaie, and Yaser Alon-achieved state-of-the-art performance among aizan. 2024. SADA: SAUDI AUDIO DATASET XLS-R Arabic models (23.03% WER), while FOR ARABIC. ICASSP, IEEE International 327 find
  2. Conference on Acoustics, Speech and Signal Processing - Proceedings, pages 10286–10290. L Alkanhal, A Alessa, E Almahmoud … of Arabic-NLP 2023, and undefined 2023. 2023. Aswat: Arabic audio dataset for automatic speech recognition using speech-representation learning. aclanthology.org, pages 120–127.
  3. NA Alrashoudi, OS Alshahri Proceedings of the 6th …, and undefined 2024. 2024. Arabic Speech Recognition of zero-resourced Languages: A Case of Shehri (Jibbali) Language. aclanthology.orgNA Alrashoudi, OS Alshahri, H Al-KhalifaProceedings of the 6th Workshop on Open-Source Arabic Corpora and, 2024•aclanthology.org, pages 84–92.
  4. Wafa Alshehri, Wasfi Al-Khatib, and Mohammad Amro. 2026. Parameter-efficient fine-tuning of xls-r for arabic speech recognition. find
  5. Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations. Advances in Neural Information Processing Systems, 33:12449–12460.
  6. M Bajo, H Fukukawa, R Morita arXiv preprint arXiv …, and undefined 2024. Efficient adaptation of multilingual models for japanese asr. arxiv.orgM Bajo, H Fukukawa, R Morita, Y OgasawaraarXiv preprint arXiv:2412.10705, 2024•arxiv.org. ↗
  7. S Bhattacharjee, J Mishra, HS Shekhawat arXiv preprint arXiv …, and undefined 2025. Parameter-Efficient Fine-Tuning of Foundation Models for CLP Speech Classification. arxiv.org. find
  8. Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, and Michael Auli. 2020. Unsupervised Cross-lingual Representation Learning for Speech Recognition. Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH, 1:346–350. find
  9. Amira Dhouib, Achraf Othman, Oussama El Ghoul, Mohamed Koutheair Khribi, Aisha Al Sinani, Amira Dhouib, Achraf Othman, Oussama El Ghoul, Mohamed Koutheair Khribi, and Aisha Al Sinani. 2022. Arabic Automatic Speech Recognition: A Systematic Literature Review. ↗
  10. Applied Sciences 2022, Vol. 12„ 12(17). find
  11. Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. 2024. Dora: Weight-decomposed low-rank adaptation. openreview.net. ↗
  12. Hüseyin Polat, Alp Kaan Turan, Cemal Koçak, Hasan Basri Ulaş, Hüseyin Polat, Alp Kaan Turan, Cemal Koçak, and Hasan Basri Ulaş. 2024. Implementation of a Whisper Architecture-Based Turkish Automatic Speech Recognition (ASR) System and Evaluation of the Effect of Fine-Tuning with a Low-Rank Adaptation (LoRA) Adapter on Its Performance. Electronics 2024, Vol. 13„ 13(21). ↗
  13. Zheshu Song, Jianheng Zhuo, Yifan Yang, Ziyang Ma, Shixiong Zhang, and Xie Chen. 2024. LoRA-Whisper: Parameter-Efficient and Extensible Multilingual ASR. Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH, pages 3934–3938. find
  14. Bashar Talafha, Abdul Waheed, and Muhammad Abdul-Mageed. 2023. N-Shot Benchmarking of Whisper on Diverse Arabic Speech Recognition. Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH, 2023-August:5092–5096.
  15. HO Toyin, A Djanibekov, A Kulkarni arXiv preprint arXiv …, and undefined 2023. ArTST: Arabic text and speech transformer. arxiv.org.
  16. HA Younis, YF Mohammad 2023 16th International, and undefined 2023. Arabic speech recognition based on self supervised learning. ieeexplore.ieee.orgHA Younis, YF Mohammad2023 16th International Conference on Developments in eSystems, 2023•ieeexplore.ieee.org. find
  17. Chin Yuen Kwok, Hexin Liu, Jia Qi Yip, Sheng Li, and Eng Siong Chng. 2025. A Two-Stage LoRA Strategy for Expanding Language Capabilities in Multilingual ASR Models. IEEE Transactions on Audio, Speech and Language Processing, 33:2576–2590. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/alshehri-2026-parameter-efficient-adaptation-self-supervised/