Elamur Research⌕
☰
home / papers / 2025 / Arabic ASR on the SADA Large-Scale Arabic Speech Corpus with Transformer-Based Models

Arabic ASR on the SADA Large-Scale Arabic Speech Corpus with Transformer-Based Models

Summary

The paper looks at how well current transformer speech recognisers handle dialectal Arabic. The test bed is SADA, a 668-hour corpus taken from 57 Saudi television shows. SADA is hard because it mixes read and spontaneous speech across several dialects and acoustic conditions, and only about a third of it counts as clean. The authors compare XLSR (El-Geish's Arabic fine-tune), XLS-R, Whisper small and medium, and MMS 1B. They then test three ways to improve them: fine-tuning on SADA, a KenLM 4-gram language model, and spectral-gating denoising with Noisereduce.

Used without adaptation, every model does badly; Whisper's word error rate even goes above 100%. Fine-tuning brings the largest gains. The best system is MMS 1B fine-tuned with an unfrozen encoder and combined with the 4-gram LM, reaching 40.9% WER and 17.6% CER on the clean test set. On the noisy subset, training on noisy audio beats cleaning the audio first, because denoising usually makes recognition worse. The work gives a realistic baseline for broadcast-style dialectal Arabic ASR on hardware within reach of most labs.

Key points

  • Data split. The environment breakdown covers roughly 437.5 h of SADA: clean 122.7 h, music 166.6 h, noisy 147.8 h and car 0.4 h. Dialects include Najdi, Hijazi and Khaleeji, plus Yemeni, Egyptian and Levantine.
  • Hardware limit. Only models that fit on an 8 GB consumer GPU were tested, so Whisper large was left out.
  • Fair comparison. All fine-tuning runs were capped at 100k steps.
  • Untuned baselines on clean test. Whisper small reached 254.9% WER and MMS 84.4% WER.
  • Fine-tuning. MMS 1B with an unfrozen encoder reached 51.5% WER and 19.9% CER.
  • Language model. The 4-gram KenLM model was trained on SADA transcripts. It is an ARPA file of about 2.2M lines and 100 MB, and it cut WER by roughly 9–11 points in absolute terms.
  • Best systems.
    • MMS_unfreeze_lm: 40.86% WER and 17.60% CER.
    • xlsr_elgeish_sada_unfreeze_lm: 42.62% WER and 18.31% CER, a good trade-off at only 300 M parameters.
  • Noise.
    • The best clean-trained XLSR scored 60.95% WER on noisy test.
    • Applying Noisereduce (coefficient 0.9) raised that to 64.46%.
    • Fine-tuning on noisy training data gave the best noisy-test result, 58.95% WER.
    • The benefit of denoising correlated weakly with baseline CER (Pearson 0.28) and barely at all with the estimated noise energy (0.03). On average it only helped once CER was above about 50%.

Interesting ideas

  • I1. Giving every model the same fixed number of training steps keeps the comparison honest, because no model wins simply by being trained longer. — For Mova (learn): When we compare parser variants or teacher tiers, train each one on an equal update budget so that differences in LAS reflect the method rather than the compute.
  • I2. The authors did not assume denoising helps; they measured how its gain changes with the difficulty of each sample and found it helps only on very hard inputs and hurts easy ones. — For Mova (learn): Self-induced repair rules could be switched on only for sentences where the witness parsers disagree or confidence is low, with a check that clean, confident cases are not made worse.
  • I3. Adapting the model to messy input worked better than cleaning the input before feeding it in. — For Mova (tree): For noisy or nonstandard English, adding such text to the gparse/perceptron training data may beat aggressive normalisation that can destroy cues the parser relies on.
  • I4. A small in-domain n-gram model placed on top of a strong neural model gave sizeable, fully inspectable gains at low cost. — For Mova (tokens): A compact count-based model trained on our own corpora could rescore tagger and lemma candidates transparently, keeping each decision explainable.
  • I5. Errors were broken down by sample length and by dialect instead of reporting a single average. — For Mova (tree): Reporting LAS sliced by sentence length, genre and source treebank would show exactly where parsing fails and which conversion of a UD dialect is responsible.

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: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2508.12968 · arXiv 2508.12968

@misc{gerazov2025arabic,
  title = {{Arabic ASR on the SADA Large-Scale Arabic Speech Corpus with Transformer-Based Models}},
  author = {Gerazov, Branislav and Politi, Marcello and Bratieres, Sebastien},
  year = {2025},
  doi = {10.48550/arxiv.2508.12968},
  eprint = {2508.12968},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2508.12968}
}

References

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

  1. 1. Abdelhamid, A.A., Alsayadi, H.A., Hegazy, I., Fayed, Z.T.: End-to-end arabic speech recognition: A review. In: Proceedings of the 19th Conference of Language Engineering. pp. 26–30 (2020) 2. Alharbi, S., Alowisheq, A., Tüske, Z., Darwish, K., Alrajeh, A., Alrowithi, A., Tamran, A.B., Ibrahim, A., Aloraini, R., Alnajim, R., et al.: Sada: Saudi audio dataset for arabic. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 10286–10290. find
  2. IEEE (2024) 3. Ardila, R., Branson, M., Davis, K., Henretty, M., Kohler, M., Meyer, J., Morais, R., Saunders, L., Tyers, F.M., Weber, G.: Common voice: A massively-multilingual speech corpus. find our card
  3. LREC (2020) 4. Babu, A., Wang, C., Tjandra, A., Lakhotia, K., Xu, Q., Goyal, N., Singh, K., Von Platen, P., Saraf, Y., Pino, J., et al.: Xls-r: Self-supervised cross-lingual speech representation learning at scale. arXiv preprint arXiv:2111.09296 (2021) 5. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33, 12449–12460 (2020) 6. Besdouri, F.Z., Zribi, I., Belguith, L.H.: Arabic automatic speech recognition: challenges and progress. Speech Communication 163, 103110 (2024) 7. Chowdhury, S.A., Hussein, A., Abdelali, A., Ali, A.: Towards one model to rule all: Multilingual strategy for dialectal code-switching Arabic ASR. arXiv preprint arXiv:2105.14779 (2021) 8. Conneau, A., Baevski, A., Collobert, R., Mohamed, A., Auli, M.: Unsupervised cross-lingual representation learning for speech recognition. arXiv preprint arXiv:2006.13979 (2020) 9. Dhouib, A., Othman, A., El Ghoul, O., Khribi, M.K., Al Sinani, A.: Arabic automatic speech recognition: A systematic literature review. Applied Sciences 12(17), 8898 (2022) 10. Gulati, A., Qin, J., Chiu, C.C., Parmar, N., Zhang, Y., Yu, J., Han, W., Wang, S., Zhang, Z., Wu, Y., et al.: Conformer: Convolution-augmented transformer for speech recognition. arXiv preprint arXiv:2005.08100 (2020) 11. Halabi, N., et al.: Arabic speech corpus. ↗ our card
  4. Oxford Text Archive Core Collection (2016) B. Gerazov et al. 12. Heafield, K.: KenLM: Faster and smaller language model queries. In: Proceedings of the Sixth Workshop on Statistical Machine Translation. pp. 187–197. Association for Computational Linguistics, Edinburgh, Scotland (Jul 2011) 13. Heafield, K., Pouzyrevsky, I., Clark, J.H., Koehn, P.: Scalable modified Kneser-Ney language model estimation. In: Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 690–696. Association for Computational Linguistics, Sofia, Bulgaria (Aug 2013) 14. Khan, H.I., Abid, A., Moussa, M.M., Abou-Allaban, A.: The Tarteel dataset: crowd-sourced and labeled Quranic recitation (2021) 15. Povey, D., Ghoshal, A., Boulianne, G., Burget, L., Glembek, O., Goel, N., Hannemann, M., Motlicek, P., Qian, Y., Schwarz, P., et al.: The kaldi speech recognition toolkit. In: IEEE 2011 workshop on automatic speech recognition and understanding. vol. 1, pp. 5–1. find our card
  5. Hawaii (2011) 16. Pratap, V., Tjandra, A., Shi, B., Tomasello, P., Babu, A., Kundu, S., Elkahky, A., Ni, Z., Vyas, A., Fazel-Zarandi, M., Baevski, A., Adi, Y., Zhang, X., Hsu, W.N., Conneau, A., Auli, M.: Scaling speech technology to 1,000+ languages. arXiv (2023) 17. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International conference on machine learning. pp. 28492–28518. find our card
  6. PMLR (2023) 18. Sainburg, T., Thielk, M., Gentner, T.Q.: Finding, visualizing, and quantifying latent structure across diverse animal vocal repertoires. PLoS computational biology 16(10), e1008228 (2020) 19. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 20. Wang, Y., Alhmoud, A., Alqurishi, M.: Open universal arabic asr leaderboard. arXiv preprint arXiv:2412.13788 (2024) 21. Young, S., Evermann, G., Gales, M., Hain, T., Kershaw, D., Liu, X., Moore, G., Odell, J., Ollason, D., Povey, D., et al.: The htk book. Cambridge university engineering department 3(175), 12 (2002) ↗ our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/gerazov-2025-arabic-asr-sada-large-scale/