Elamur Research⌕
☰
home / papers / 2026 / Eraserhead at OSACT7 Shared Task: ASR Consistency Filtering and Speaker-Adaptive Post-Processing for Arabic Speech Diacritization

Eraserhead at OSACT7 Shared Task: ASR Consistency Filtering and Speaker-Adaptive Post-Processing for Arabic Speech Diacritization

Paper · 2026 · Muhammad Abu Horaira and Nahian Chowdhury

Summary

This paper describes a 4th-place system for Task 2 of the KSAA-2026 Shared Task. The task gives a system Arabic speech plus its undiacritized transcript and asks it to restore all diacritics character by character. Noisy ASR output, dialect variety and speaker habits around word-final case endings make this hard. The authors start from the pretrained Diac Transformer, which fuses a text branch with a Whisper-transcript branch through cross-attention, and add four pieces:

  • ASR Consistency Filtering: a low-agreement ASR transcript is swapped for the undiacritized gold text rather than being dropped.
  • Checkpoint ensembling: three checkpoints are combined by per-character confidence voting.
  • Speaker-adaptive rewriting: word-final endings are changed to sukun for speakers who habitually drop case vowels.
  • Constrained decoding: a final guard against outputs whose letters differ from the input.

Training used only the official corpus of about five hours. On the test set the system reached 8.23 DER and 30.37 WER (WCE, Incl. 0). Out of four evaluation settings, it outperformed the Text+ASR baseline that the organizers had fine-tuned in all but one. The work shows that cheap data-side and output-side heuristics can matter as much as architecture in low-resource settings.

Key points

  • Data: the KSAA-2026 corpus covers MSA and dialectal speech from nine nationalities, split into 2,327 train, 260 dev and 260 test utterances. Clips are 16 kHz and at most 9 s long, and no external data was allowed.

  • Models: the base is rufaelfekadu/diac-transformer-text-asr-tashkeela-clartts (about 1.63M parameters, d_model 256, two blocks). ASR transcripts came from a frozen sashat/whisper-medium-ClassicalAr.

  • Filtering: agreement is a SequenceMatcher ratio between gold and ASR text, with spaces and diacritics removed. The softer threshold τ=0.5 beat τ=0.9 on dev (32.75 vs 33.38 WER).

  • Dev ablation (WER):

    StageWER
    Pretrained baseline43.07
    Fine-tuned, τ=0.532.75
    + ensemble32.15
    + speaker post-processing (θ=0.65 from grid search)31.23

  • Test results: DER/WER/SER were 8.23/30.37/80.79, against 9.91/31.84/82.93 for the baseline. The system lost to the baseline only in the Excl. 0, WOCE condition (DER 5.87 vs 4.82).

Interesting ideas

  • I1. When an auxiliary input looks unreliable, replace it with a neutral stand-in derived from gold, never the target labels; this keeps the example and avoids leaking answers
  • I2. Ensemble by letting the most confident model win at each position, instead of averaging probabilities; every decision can be traced to one named model
  • I3. Build per-source style profiles on dev data, act only on sources with a strong preference, and treat unseen sources as neutral
  • I4. A cheap consistency guard reverts any output whose base letters no longer match the input token
  • I5. Tune thresholds on the strict whole-unit metric (WER) rather than the per-character one, since partial correctness can hide errors that break the whole unit

For Mova

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

  • I1 (learn): when witness parsers disagree on a feature for a silver sentence, feed a neutral or gold-derived placeholder instead of discarding the sentence, so the training set keeps its coverage.
  • I2 (tokens): combine several taggers or lemmatizers token by token this way, and log which model won each token so the choice stays explainable.
  • I3 (tree): keep per-treebank or per-genre profiles of annotation habits, and apply dialect-conversion rewrites only where the profile is decisive.
  • I4 (back): in the text → graph → text compressor, check that regenerated content words match the source lemmas, and fall back to the original token when they do not.
  • I5 (learn): select repair-rule and parser thresholds by whole-tree or whole-frame exact match rather than per-arc scores.

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

🎓Chittagong University of Engineering & TechnologyChittagong · Bangladesh25 works · 29 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 worksComputer Vision and Pattern Recognition · 2,869 works

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

Cite

Cite as Muhammad Abu Horaira and Nahian Chowdhury. 2026. Eraserhead at OSACT7 Shared Task: ASR Consistency Filtering and Speaker-Adaptive Post-Processing for Arabic Speech Diacritization. The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/4x2nvon8vija

@inproceedings{horaira2026eraserhead,
  title = {{Eraserhead at OSACT7 Shared Task: ASR Consistency Filtering and Speaker-Adaptive Post-Processing for Arabic Speech Diacritization}},
  author = {Muhammad Abu Horaira and Nahian Chowdhury},
  booktitle = {The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks},
  year = {2026},
  doi = {10.63317/4x2nvon8vija},
  url = {https://doi.org/10.63317/4x2nvon8vija}
}

References

11 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Asma Al Wazrah, Waad Alshammari, Rawan Al-matham, Raghad Al-Rasheed, Afrah Altamimi, Rufael Marew, Sawsan Alqahtani, Hanan Aldarmaki, Abdullah Alharbi, Abdulrahman Alshehri, Mohamed Assar, Amal Almazrua, and Abdulrahman AlOsaimy. 2026. Ksaa-2026 shared task on arabic speech dictation with automatic diacritization. In Proceedings of the 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7).
  2. Hanan Aldarmaki and Ahmad Ghannam. 2023. Diacritic recognition performance in arabic asr. pages 361–365.
  3. Badr AlKhamissi, Muhammad ElNokrashy, and Mohamed Gabr. 2020. Deep diacritization: Efficient hierarchical recurrence for improved Arabic diacritization. In Proceedings of the Fifth Arabic Natural Language Processing Workshop, pages 38–48, Barcelona, Spain (Online). Association for Computational Linguistics. ↗
  4. Salman Elgamal, Ossama Obeid, Mhd Kabbani, Go Inoue, and Nizar Habash. 2024. Arabic diacritics in the wild: Exploiting opportunities for improved diacritization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14815–14829, Bangkok, Thailand. Association for Computational Linguistics.
  5. Ali Fadel, Ibraheem Tuffaha, Bara’ Al-Jawarneh, and Mahmoud Al-Ayyoub. 2019. Neural Arabic text diacritization: State of the art results and a novel approach for machine translation. In Proceedings of the 6th Workshop on Asian Translation, pages 215–225, Hong Kong, China. Association for Computational Linguistics. ↗
  6. Hamdy Mubarak, Ahmed Abdelali, Hassan Sajjad, Younes Samih, and Kareem Darwish. 2019. Highly effective Arabic diacritization using sequence to sequence modeling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2390– 2395, Minneapolis, Minnesota. Association for Computational Linguistics.
  7. Han Qin, Guimin Chen, Yuanhe Tian, and Yan Song. 2021. Improving Arabic diacritization with regularized decoding and adversarial training. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 534–542, Online. Association for Computational Linguistics.
  8. Sara Shatnawi, Sawsan Alqahtani, and Hanan Aldarmaki. 2024. Automatic restoration of diacritics for speech data sets. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 4166–4176, Mexico City, Mexico. Association for Computational Linguistics. ↗
  9. Brian Thompson and Ali Alshehri. 2022. Improving Arabic diacritization by learning to diacritize and translate. In Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 11–21, Dublin, Ireland (in-person and online). Association for Computational Linguistics. 8.
  10. Language Resource References Fekadu, Rufael. 2024. Diac Transformer: Text and ASR-based Arabic Diacritization Model. Hugging Face model repository. find
  11. Shatnawi, Sara. 2024. Whisper Medium Classical Arabic. Hugging Face model repository. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/horaira-2026-eraserhead-osact7-shared-task-asr/