Elamur Research⌕
☰
home / papers / 2026 / Inverse Text Normalization for Arabic Numbers in Streaming ASR

Inverse Text Normalization for Arabic Numbers in Streaming ASR

Paper · 2026 · Enas Albasiri, Myungjong Kim, Nourchene Ferchichi et al.

Summary

The paper asks whether a streaming speech recognizer can write Arabic numbers directly as digits, so that no separate inverse text normalization (ITN) stage is needed after decoding. Arabic makes this hard. Numeral forms depend on the counted noun through gender polarity and case. Dialects verbalize numbers differently, spelling is inconsistent, and diacritics may or may not appear.

The authors fine-tuned two models from the same 600M-parameter multilingual cache-aware FastConformer:

  • Baseline: trained on spoken-form transcripts.
  • Unified: trained on the same transcripts after they were converted to written form by NeMo's WFST-based ITN grammar.

Training used about 7600 hours of Arabic (MSA, dialectal, classical) and roughly 80,000 hours of English. Training on written-form targets left recognition accuracy essentially unchanged. The unified model also matched, and slightly beat, the cascade of baseline plus rule-based ITN on number exact match.

An analysis of the unified model's Arabic errors found that most failures come from diacritized output, wrong gender marking and spelling variants. This shows where end-to-end numeral handling for morphologically rich languages still breaks down.

Key points

  • Architecture: cache-aware streaming FastConformer with a hybrid CTC/RNNT decoder, pretrained on 40 languages with a universal BPE tokenizer.
    • Diacritics are separate output tokens.
    • Inference uses RNNT with a 0.56 s chunk size.
  • Arabic ASR: average WER was 44.62% for the unified model against 45.11% for the baseline, on Casablanca, Common Voice, MGB-2 and SADA.
  • English ASR: average WER rose slightly, from 6.67% to 7.00%, on LibriSpeech, Tedlium2 and SPGISpeech.
  • ITN exact match:
    • Arabic, 2,768 items: unified 74.64% vs cascade 74.31%.
    • English, 4,247 items built from the Google text normalization corpus via MagpieTTS: unified 82.83% vs cascade 81.28%.
  • Error analysis of the 702 unified Arabic errors:
    • 66.7% left the number in spoken form instead of digits.
    • By category: diacritics 27.9%, gender mismatch 22.7%, partial conversion 11.1%, orthographic variation 10.8%, missing number 9.1%.
  • Rule-based ITN grammar (Pynini, NeMo):
    • It is context-free and MSA-only.
    • It defaults to feminine gender and nominative case.
    • It covers only cardinals, decimals, fractions and money.

Interesting ideas

  • I1. To compare a digit-producing model fairly with spoken-form references, the authors pass its output back through text normalization before scoring WER, so the metric measures recognition rather than format.
  • I2. The authors openly note that their gold written forms come from a deterministic grammar with fixed defaults, so the gold reflects one convention, not the only correct answer.
  • I3. Gender polarity shows that a numeral's correct form depends on its head noun, so a normalizer that sees only the number token is bound to fail.
  • I4. They propose a hybrid: deterministic transducers handle confident cases, and an LLM handles only context-dependent ones, to produce cleaner training targets while avoiding plausible-but-wrong LLM outputs.
  • I5. Errors are broken down into named categories with percentages, which turns one accuracy number into a ranked list of fixes.

For Mova

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

  • I1 (learn): when comparing parsers or generators trained on different surface conventions, map outputs to one canonical form before scoring, so convention differences are not counted as errors.
  • I2 (learn): silver data from rule-based converters should be tagged with the conventions it bakes in, so evaluation separates real mistakes from mere disagreements with those conventions.
  • I3 (world): resolve quantities through the nummod link to the counted noun in the UD tree, so value, unit and agreement are read together rather than from the numeral alone.
  • I4 (learn): route only low-confidence or context-dependent items to LLM teacher tiers and keep rule outputs for the rest, which limits unrecoverable teacher errors in silver data.
  • I5 (answer): classify failures on school tests or word problems (e.g., misread quantity, wrong unit, missed entity) to decide which earlier stage to repair first.

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

🏢NvidiaSanta Clara · United States25 works · 5,058 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 Enas Albasiri, Myungjong Kim, Nourchene Ferchichi and Oluwatobi Olabiyi. 2026. Inverse Text Normalization for Arabic Numbers in Streaming ASR. Proceedings of the Third Workshop on Computation and Written Language (CAWL 2026) @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/4wy5qhn7npqa

@inproceedings{albasiri2026inverse,
  title = {{Inverse Text Normalization for Arabic Numbers in Streaming ASR}},
  author = {Enas Albasiri, Myungjong Kim, Nourchene Ferchichi and Oluwatobi Olabiyi},
  booktitle = {Proceedings of the Third Workshop on Computation and Written Language (CAWL 2026) @ LREC 2026},
  year = {2026},
  doi = {10.63317/4wy5qhn7npqa},
  url = {https://doi.org/10.63317/4wy5qhn7npqa}
}

References

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

  1. Imad Al-Sughaiyer and Ibrahim Al-Kharashi. 2004. Arabic morphological analysis techniques: A comprehensive survey. JASIST, 55:189–213. ↗
  2. Sarah Alkuhlani and Nizar Habash. 2011. A corpus for modeling morpho-syntactic agreement in Arabic: Gender, number and rationality. In ACL-HLT 2011 - Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 357–362.
  3. Ahmad Alqassas. 2017. Gender and number polarity in modern standard arabic numeral phrases. Canadian Journal of Linguistics/Revue canadienne de linguistique, 62(1):1–17. ↗
  4. Dario Amodei, Rishita Anubhai, Eric Battenberg, Carl Case, Jared Casper, Bryan Catanzaro, et al. 2015. Deep speech 2: End-to-end speech recognition in english and mandarin. ↗
  5. Suzanne Mahmoud Bardeas. 2009. The Syntax of the Arabic DP. Ph.D. thesis, University of York. find
  6. Ahmed El-kholy and Nizar Habash. 2011. Automatic error analysis for morphologically rich languages. find
  7. Hassan Gadalla. 2023. Comparative Morphology of Standard and Egyptian Arabic. find
  8. Hassan Abdel-Shafik Hassan Gadalla. 2000. Numerals in standard arabic and egyptian colloquial arabic: A comparative study. Bulletin of the Faculty of Arts, Assiut University, 4(2):149– 183. find
  9. Yashesh Gaur, Nick Kibre, Jian Xue, Kangyuan Shu, Yuhui Wang, Issac Alphanso, Jinyu Li, and Yifan Gong. 2023. Streaming, fast and accurate on-device inverse text normalization for automatic speech recognition. In IEEE Spoken Language Technology Workshop (SLT). IEEE. find
  10. Kyle Gorman. 2016. Pynini: A Python library for weighted finite-state grammar compilation. In Proceedings of the SIGFSM Workshop on Statistical NLP and Weighted Automata, pages 75– 80, Berlin, Germany. Association for Computational Linguistics. find
  11. Kyle Gorman and Richard Sproat. 2016. Minimally supervised number normalization. Transactions of the Association for Computational Linguistics, 4:507–519. ↗
  12. Nizar Habash, Ryan Roth, Owen Rambow, Ramy Eskander, and Nadi Tomeh. 2013. Morphological analysis and disambiguation for dialectal arabic. In Proceedings of NAACL-HLT 2013.
  13. Nizar Y. Habash. 2010. Introduction to Arabic Natural Language Processing. Synthesis Lectures on Human Language Technologies. Morgan and Claypool Publishers. ↗
  14. Go Inoue, Salam Khalifa, and Nizar Habash. 2022. Morphosyntactic tagging with pre-trained language models for Arabic and its dialects. In Findings of the Association for Computational Linguistics: ACL 2022, pages 1708–1719, Dublin, Ireland. Association for Computational Linguistics.
  15. Paarth Neekhara, Jason Huang, Kushal Lakhotia, Vitaly Lavrukhin, Boris Ginsburg, and Jagadeesh Balam. 2024. Improving robustness of llm-based speech synthesis by learning monotonic alignment. In Interspeech 2024. find
  16. Thai Binh Nguyen, Xin Zhang, Yufeng Li, Yashesh Gaur, and Yifan Gong. 2023. Four-in-one: A joint single-pass model for inverse text normalization, punctuation, capitalization, and disfluency. In INTERSPEECH 2023. find
  17. Vahid Noroozi, Somshubra Majumdar, Ankur Kumar, Jagadeesh Balam, and Boris Ginsburg. 2024. Stateful conformer with cache-based inference for streaming automatic speech recognition. In ICASSP 2024. IEEE. find
  18. Dima Rekesh, Nithin Rao Koluguri, Samuel Kriman, Somshubra Majumdar, Vahid Noroozi, He Huang, Oleksii Hrinchuk, Krishna C. Puvvada, Ankur Kumar, Jagadeesh Balam, and Boris Ginsburg. 2023. Fast conformer with linearly scalable attention for efficient speech recognition. In ASRU, pages 1–8. IEEE. find
  19. Julian Salazar et al. 2024. Handling numeric expressions in automatic speech recognition. find
  20. George Saon, Gakuto Kurata, Tom Sercu, Kartik Audhkhasi, Samuel Thomas, Dimitrios Dimitriadis, Xiaodong Cui, Bhuvana Ramabhadran, Michael Picheny, Lynn-Li Lim, Bergul Roomi, and Phil Hall. 2017. English conversational telephone speech recognition by humans and machines.
  21. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715–1725, Berlin, Germany. Association for Computational Linguistics. ↗
  22. Abdelhadi Soudi, Günter Neumann, and Antal Van den Bosch. 2007. Arabic Computational Morphology: Knowledge-based and Empirical Methods, volume 38, pages 3–14. find
  23. Richard Sproat and Navdeep Jaitly. 2017. Rnn approaches to text normalization: A challenge. ↗
  24. Zhong-Qiu Tang et al. 2025. Universal-2-tf: Robust all-neural text formatting for asr. find
  25. Yingzhi Wang, Anas Alhmoud, and Muhammad Alqurishi. 2024. Open universal arabic asr leaderboard.
  26. W. Xiong, J. Droppo, X. Huang, F. Seide, M. Seltzer, A. Stolcke, D. Yu, and G. Zweig. 2017. Achieving human parity in conversational speech recognition. ↗
  27. Yang Zhang, Evelina Bakhturina, Kyle Gorman, and Boris Ginsburg. 2021. Nemo inverse text normalization: From development to production. ↗
  28. Yang Zhang, Travis M. Bartley, Mariana Graterol-Fuenmayor, Vitaly Lavrukhin, Evelina Bakhturina, and Boris Ginsburg. 2023. A chat about boring problems: Studying gpt-based text normalization. CoRR, abs/2309.13426. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/albasiri-2026-inverse-text-normalization-arabic-numbers/