---
title: "DialG2P: Dialectal Grapheme-to-Phoneme. Arabic as a Case Study"
description: "The paper addresses grapheme-to-phoneme (G2P) conversion for dialectal Arabic, using Egyptian (EGY) as the… 5.15% WER. 25 references."
authors: ["Majd Hawasly", "Hamdy Mubarak", "Ahmed Abdelali", "Ahmed Ali"]
year: 2025
venue: "Proceedings of The Third Arabic Natural Language Processing Conference"
source: "https://aclanthology.org/2025.arabicnlp-main.38/"
doi: "10.18653/v1/2025.arabicnlp-main.38"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/hawasly-2025-dialg2p-dialectal-grapheme-phoneme-arabic/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/hawasly-2025-dialg2p-dialectal-grapheme-phoneme-arabic/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/14/) / DialG2P: Dialectal Grapheme-to-Phoneme. Arabic as a Case Study

# DialG2P: Dialectal Grapheme-to-Phoneme. Arabic as a Case Study

**Authors:** [Majd Hawasly](https://md.research.elamur.ai/free/en/authors/majd-hawasly/), [Hamdy Mubarak](https://md.research.elamur.ai/free/en/authors/hamdy-mubarak/), [Ahmed Abdelali](https://md.research.elamur.ai/free/en/authors/ahmed-abdelali/) and [Ahmed Ali](https://md.research.elamur.ai/free/en/authors/ahmed-m-abdelatty-ali/) · **Year:** 2025 · **Venue:** [Proceedings of The Third Arabic Natural Language Processing Conference](https://md.research.elamur.ai/free/en/venues/arabicnlp/) ([ArabicNLP 2025](https://md.research.elamur.ai/free/en/venues/arabicnlp/2025/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2025.arabicnlp-main.38/) · [PDF](https://aclanthology.org/2025.arabicnlp-main.38.pdf) · [DOI 10.18653/v1/2025.arabicnlp-main.38](https://doi.org/10.18653/v1/2025.arabicnlp-main.38) · [OpenAlex](https://openalex.org/W4416037649)

**Citations:** 1 citations (OpenAlex, 2026-10-14)

## Summary
The paper addresses grapheme-to-phoneme (G2P) conversion for dialectal Arabic, using Egyptian (EGY) as the test case. EGY has no standard spelling, and tweets are usually written without short vowels. Many letters are also pronounced differently than in MSA, and some borrowed sounds (g, v, p) have no native letter. The authors built a resource from 10,000 QADI tweets. A native-speaker linguist fully diacritized them and recorded the verbatim Cairene pronunciation. A comparable Levantine (LEV) sample was annotated for comparison. Their model, DialG2P, is a single character-level seq2seq Transformer trained on two tasks at once: restoring diacritics and replacing letters with dialect-specific "special sound" symbols. A fixed 1-to-1 table then turns the output characters into phonemes. DialG2P reaches 5.15% WER and 1.71% PER. This is roughly equal to a cascade of a diacritizer plus rules, and far better than GPT-4 in zero-shot or 10-shot settings. The results show that dialectal G2P needs more than diacritization. They also suggest that one unified network can absorb the dialect-specific rewrites.

## Key points
- **Data:** 10,000 EGY tweets from the QADI corpus, diacritized with verbatim Cairo pronunciation by a single annotator, split 80/10/10. A further 10,000 LEV tweets were annotated by a Syrian speaker. The diacritized data is released at github.com/qcri/DialG2P.
- **Special-sound statistics:**
  - j→g occurs 98.5% of the time in EGY but only 2.46% in LEV.
  - q→glottal stop is the most common shared shift, at 80.56% in EGY and 77.37% in LEV.
- **Model and training:**
  - Encoder-decoder Transformer with 6 layers, 512 hidden units and 8 heads.
  - Trained for 300,000 steps with batch size 512 and LazyAdam.
  - Input is 7-word sliding windows, padded with six start and six end tokens and split into letters.
- **Results on the 1,000-tweet test set:**

  | System | WER | PER | DER | CER |
  |---|---|---|---|---|
  | DialG2P | 5.15 | 1.71 | 1.67 | 0.05 (best) |
  | Transformer cascade | 5.11 | 1.70 | 1.62 | 0.09 |
  | Diacritization-only Transformer | 17.26 | – | – | – |
  | Diacritization-only Transformer + defSS | 6.32 | – | – | – |
  | GPT-4 10-shot + defSS | 25.14 | – | – | – |
  | GPT-4 0-shot (no defSS) | 47.57 | – | – | – |

- **Stated limitations:** only one dialect was evaluated, there was a single annotator, and the gold letter substitutions were produced by a rule base. That last point partly favours the rule-based cascade.

## Interesting ideas
- Every word appears in seven overlapping windows, and a per-letter majority vote picks the final output, so each word gets a small ensemble of contexts at no extra model cost.
- The second task is built into the output alphabet by adding special symbols for dialect sounds, so one decoder does both diacritization and substitution without a separate head.
- A cheap "default special sounds" fix, which applies only substitutions that happen in roughly 80% or more of cases, greatly improved both the plain Transformer and GPT-4 outputs.
- Errors are reported at four granularities (word, phoneme, diacritic, character), which shows whether a system fails at vowel restoration or at sound substitution.
- Corpus frequencies of each sound shift per dialect are used to decide which rewrites become defaults and which stay lexical exceptions.

## For Mova

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

- **I1** (tokens): The tagger and morphology could be run over overlapping sentence windows, voting per token and keeping the vote split as an explainable confidence signal.
- **I2** (tokens): A lemmatizer could output a normalized form with inline markers for spelling variants or contractions, giving one joint decision instead of a brittle pipeline.
- **I3** (learn): LLM teacher outputs could pass through a deterministic majority-rule correction step before becoming silver data, with the improvement measured separately so the rules stay auditable.
- **I4** (learn): Evaluation could be split by layer (token, arc, label, frame role), so a regression is traced to the stage that caused it rather than only to an end score.
- **I5** (tree): UD dialect conversion could count how often each mapping holds in the treebank, turning high-frequency ones into default rules and flagging the rest for case-by-case handling.

*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

- [Hamad bin Khalifa University](https://md.research.elamur.ai/free/en/universities/hamad-bin-khalifa-university/) — University, Doha, Qatar · 141 works · 14,175 citations

📚 Shelves (6): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [💭 Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/) · 1,552 works · [💭 Experimental and Cognitive Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/experimental-and-cognitive-psychology/) · 870 works

🧭 Topics (1): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works

## Cite

Cite as: Majd Hawasly, Hamdy Mubarak, Ahmed Abdelali and Ahmed Ali. 2025. DialG2P: Dialectal Grapheme-to-Phoneme. Arabic as a Case Study. Proceedings of The Third Arabic Natural Language Processing Conference.

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

```bibtex
@inproceedings{hawasly2025dialg2p,
  title = {{DialG2P: Dialectal Grapheme-to-Phoneme. Arabic as a Case Study}},
  author = {Majd Hawasly, Hamdy Mubarak, Ahmed Abdelali and Ahmed Ali},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-main.38},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-main.38}
}
```

- [← previous Development of Android Game-Based Media to Enhance Students' Arabic Grammar Skills](https://md.research.elamur.ai/free/en/papers/utomo-2025-development-android-game-based-media/)
- [next → Egyhealth at General Arabic Health QA (MedArabiQ): An Enhanced RAG Framework with Large-Scale Arabic Q&A Medical Data](https://md.research.elamur.ai/free/en/papers/amer-2025-egyhealth-general-arabic-health-qa/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/14/)

## References

25 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Ahmed Abdelali, Hamdy Mubarak, Younes Samih, Sabit Hassan, and Kareem Darwish. 2021. [QADI: Arabic dialect identification in the wild](https://md.research.elamur.ai/free/en/papers/abdelali-2021-qadi-arabic-dialect-identification-wild/). In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 1–10.
2. Salman H Al-Ani. 1970. Arabic phonology: An acoustical and physiological Investigation. The Hague, Mouton. [find](https://scholar.google.com/scholar?q=Salman+H+Al%2DAni%2E+1970%2E+Arabic+phonology%3A+An+acoustical+and+physiological+Investigation%2E+The+Hague%2C+Mouton%2E)
3. Fawaz S. Al-Anzi and Dia Abuzeina. 2017. The impact of phonological rules on Arabic speech recognition. Int. J. Speech Technol., 20(3):715–723. [↗](https://doi.org/10.1007/s10772-017-9440-2)
4. Salima Harrat, Mourad Abbas, Karima Meftouh, and Kamel Smaili. 2013. Diacritics restoration for Arabic dialects. In INTERSPEECH 2013-14th Annual Conference of the International Speech Communication Association. [find](https://scholar.google.com/scholar?q=Salima+Harrat%2C+Mourad+Abbas%2C+Karima+Meftouh%2C+and+Kamel+Smaili%2E+2013%2E+Diacritics+restoration+for+Arabic+dialects%2E+In+INTERSPEECH+2013%2D14th+Annual+Conference+of+the+International+Speech+Communication+As)
5. Salima Harrat, Karima Meftouh, Mourad Abbas, and Kamel Smaïli. 2014. Grapheme to phoneme conversion-an Arabic dialect case. In Spoken Language Technologies for Under-resourced Languages. [find](https://scholar.google.com/scholar?q=Salima+Harrat%2C+Karima+Meftouh%2C+Mourad+Abbas%2C+and+Kamel+Sma%C3%AFli%2E+2014%2E+Grapheme+to+phoneme+conversion%2Dan+Arabic+dialect+case%2E+In+Spoken+Language+Technologies+for+Under%2Dresourced+Languages%2E)
6. Sittichai Jiampojamarn and Grzegorz Kondrak. 2010. Letter-phoneme alignment: An exploration. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 780–788, Uppsala, Sweden. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Sittichai+Jiampojamarn+and+Grzegorz+Kondrak%2E+2010%2E+Letter%2Dphoneme+alignment%3A+An+exploration%2E+In+Proceedings+of+the+48th+Annual+Meeting+of+the+Association+for+Computational+Linguistics%2C+pages+780%E2%80%93788%2C+)
7. Ronald M. Kaplan and Martin Kay. 1994. Regular models of phonological rule systems. Computational Linguistics, 20(3):331–378. [↗](https://doi.org/10.5555/204915.204917)
8. Hassan Al-Haj, Roger Hsiao, Ian Lane, Alan W Black, and Alex Waibel. 2009. Pronunciation modeling for dialectal Arabic speech recognition. In 2009 IEEE Workshop on Automatic Speech Recognition & Understanding, pages 525–528. IEEE. [↗](https://doi.org/10.5445/ir/1000166347)
9. Xinjian Li, Florian Metze, David R Mortensen, Shinji Watanabe, and Alan W Black. 2022. Zero-shot learning for grapheme to phoneme conversion with language ensemble. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2106– 2115. [find](https://scholar.google.com/scholar?q=Xinjian+Li%2C+Florian+Metze%2C+David+R+Mortensen%2C+Shinji+Watanabe%2C+and+Alan+W+Black%2E+2022%2E+Zero%2Dshot+learning+for+grapheme+to+phoneme+conversion+with+language+ensemble%2E+In+Findings+of+the+Association+for+)
10. Mansour Alghamdi, Husni Almuhtasib, and Mustafa Elshafei. 2004. Arabic phonological rules. King Saud University Journal: Computer Sciences and Information, 16:1–25. [find](https://scholar.google.com/scholar?q=Mansour+Alghamdi%2C+Husni+Almuhtasib%2C+and+Mustafa+Elshafei%2E+2004%2E+Arabic+phonological+rules%2E+King+Saud+University+Journal%3A+Computer+Sciences+and+Information%2C+16%3A1%E2%80%9325%2E)
11. Ahmed Ali, Yifan Zhang, Patrick Cardinal, Najim Dahak, Stephan Vogel, and James Glass. 2014. A complete KALDI recipe for building Arabic speech recognition systems. In 2014 IEEE spoken language technology workshop (SLT), pages 525–529. IEEE. [find](https://scholar.google.com/scholar?q=Ahmed+Ali%2C+Yifan+Zhang%2C+Patrick+Cardinal%2C+Najim+Dahak%2C+Stephan+Vogel%2C+and+James+Glass%2E+2014%2E+A+complete+KALDI+recipe+for+building+Arabic+speech+recognition+systems%2E+In+2014+IEEE+spoken+language+techno)
12. Fadi Biadsy, Nizar Habash, and Julia Hirschberg. 2009. [Improving the Arabic pronunciation dictionary for phone and word recognition with linguistically-based pronunciation rules](https://md.research.elamur.ai/free/en/papers/biadsy-2009-improving-arabic-pronunciation-dictionary-phone/). In Proceedings of human language technologies: The 2009 annual conference of the North American chapter of the association for computational linguistics, pages 397–405.
13. Maximilian Bisani and Hermann Ney. 2008. Joint-sequence models for grapheme-to-phoneme conversion. Speech communication, 50(5):434–451. [↗](https://doi.org/10.1016/j.specom.2008.01.002)
14. Ryan Cotterell and Chris Callison-Burch. 2014. [A multi-dialect, multi-genre corpus of informal written Arabic](https://md.research.elamur.ai/free/en/papers/cotterell-2014-multi-dialect-multi-genre-corpus/). In LREC, pages 241–245.
15. Hamdy Mubarak, Ahmed Abdelali, Kareem Darwish, Mohamed Eldesouki, Younes Samih, and Hassan Sajjad. 2019a. [A system for diacritizing four varieties of Arabic](https://md.research.elamur.ai/free/en/papers/mubarak-2019-system-diacritizing-four-varieties-arabic/). In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP): System Demonstrations, pages 217–222.
16. Hamdy Mubarak, Ahmed Abdelali, Hassan Sajjad, Younes Samih, and Kareem Darwish. 2019b. [Highly effective Arabic diacritization using sequence to sequence modeling](https://md.research.elamur.ai/free/en/papers/mubarak-2019-highly-effective-arabic-diacritization-sequence/). 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.
17. Hamdy Mubarak and Kareem Darwish. 2014. [Using twitter to collect a multi-dialectal corpus of Arabic](https://md.research.elamur.ai/free/en/papers/mubarak-2014-twitter-collect-multi-dialectal-corpus/). In Proceedings of the EMNLP 2014 Workshop on Arabic Natural Language Processing (ANLP), pages 1–7.
18. Ben Peters, Jon Dehdari, and Josef van Genabith. 2017. Massively multilingual neural grapheme-to-phoneme conversion. arXiv preprint arXiv:1708.01464. [↗](https://arxiv.org/abs/1708.01464)
19. Kanishka Rao, Fuchun Peng, Haşim Sak, and Françoise Beaufays. 2015. Grapheme-to-phoneme conversion using long short-term memory recurrent neural networks. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4225–4229. [find](https://scholar.google.com/scholar?q=Kanishka+Rao%2C+Fuchun+Peng%2C+Has%CC%A7im+Sak%2C+and+Fran%C3%A7oise+Beaufays%2E+2015%2E+Grapheme%2Dto%2Dphoneme+conversion+using+long+short%2Dterm+memory+recurrent+neural+networks%2E+In+2015+IEEE+International+Conference+on+Aco)
20. Pamela M Rogerson-Revell. 2021. Computer-assisted pronunciation training (capt): Current issues and future directions. Relc Journal, 52(1):189–205. [↗](https://doi.org/10.1177/0033688220977406)
21. Andreas Stolcke. 2002. SRILM-an extensible language modeling toolkit. In Seventh international conference on spoken language processing. [find](https://scholar.google.com/scholar?q=Andreas+Stolcke%2E+2002%2E+SRILM%2Dan+extensible+language+modeling+toolkit%2E+In+Seventh+international+conference+on+spoken+language+processing%2E)
22. Xu Tan, Tao Qin, Frank Soong, and Tie-Yan Liu. 2021. A survey on neural speech synthesis. arXiv preprint arXiv:2106.15561. [↗](https://arxiv.org/abs/2106.15561)
23. TensorFlow. 2019. Lazyadamoptimizer. [https://github.com/tensorflow/tensorflow/](https://github.com/tensorflow/tensorflow/) blob/r1.13/tensorflow/contrib/opt/python/ training/lazy_adam_optimizer.py. Accessed: 2025-07-06. [↗](https://github.com/tensorflow/tensorflow/)
24. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, L ukasz Kaiser, and Illia Polosukhin. 2017. [Attention is all you need](https://md.research.elamur.ai/free/en/papers/vaswani-2017-attention-all-need/). In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc.
25. Kaisheng Yao and Geoffrey Zweig. 2015. Sequence-to-sequence neural net models for grapheme-to-phoneme conversion. arXiv preprint arXiv:1506.00196. [↗](https://arxiv.org/abs/1506.00196)

*Updated 1 November 2026*