---
title: "Unsupervised Method for Improving Arabic Speech Recognition"
description: "The authors propose an unsupervised post-correction step that works on the text output. 15 references."
authors: ["Mohamed Labidi", "Mohsen Maraoui", "Mounir Zrigui"]
year: 2017
venue: "Proceedings of the 31st Pacific Asia Conference on Language, Information and Computation"
source: "https://aclanthology.org/Y17-1024/"
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/labidi-2017-unsupervised-method-improving-arabic-speech/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/labidi-2017-unsupervised-method-improving-arabic-speech/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2017](https://md.research.elamur.ai/free/en/papers/2017/page/6/) / Unsupervised Method for Improving Arabic Speech Recognition Systems

# Unsupervised Method for Improving Arabic Speech Recognition Systems

**Authors:** Mohamed Labidi, [Mohsen Maraoui](https://md.research.elamur.ai/free/en/authors/mohsen-maraoui/) and [Mounir Zrigui](https://md.research.elamur.ai/free/en/authors/mounir-zrigui/) · **Year:** 2017 · **Venue:** [Proceedings of the 31st Pacific Asia Conference on Language, Information and Computation](https://md.research.elamur.ai/free/en/venues/paclic/) ([PACLIC 2017](https://md.research.elamur.ai/free/en/venues/paclic/2017/)) · **Publisher:** [The National University (Phillippines)](https://md.research.elamur.ai/free/en/publishers/the-national-university-phillippines/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/Y17-1024/) · [PDF](https://aclanthology.org/Y17-1024.pdf) · [OpenAlex](https://openalex.org/W2888882791)

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

## Summary
The paper addresses out-of-vocabulary (OOV) words and misrecognitions in large-vocabulary Arabic ASR. Retraining the language model whenever new words appear is costly, and re-decoding audio is usually impractical. The authors propose an unsupervised post-correction step that works on the text output. A CMU Sphinx recognizer is trained on 51h of Modern Standard Arabic, with an SRILM language model built from roughly 200M words. A separate web crawler keeps gathering Arabic text and builds a directed weighted graph of word succession, which can be updated without retraining. Any transcribed word missing from the graph is flagged as an error. Context windows of growing width around the error are then used to look up candidate replacements in the graph, and the window size that yields the fewest candidates is kept. The output keeps all surviving candidates, each with its succession probabilities. On an 8h test set, WER falls from 12.5% to 8.11%, and to 7.9% after a graph update. The paper calls this a 4.6% reduction; it is actually 4.6 absolute points. The authors claim it beats earlier Arabic OOV methods.

## Key points
- **Recognizer and language model:** CMU Sphinx acoustic model trained on 51h of MSA speech from 41 native speakers (16 kHz, 16-bit, mono WAV). The SRILM language model uses about 200M words from Ajdir, Tashkeela, Abbas, OSAC and crawled corpora.
- **Test set:** 8h split into 12 files of about 40 minutes each, totalling around 48,000 words, of which 2,000 are OOV.
- **Crawler and graph:**
  - Sites are found through search-engine APIs and filtered for trustworthiness with the WOT tool. The first crawl produced over 2,981 text files.
  - Each graph node stores the word, its frequency, its date of first use, links to following words, and arc weights (bigram-style succession probabilities).
- **Results:**
  - Errors dropped from 6,000 to 3,896 after the first pass. Of the 2,104 corrected words, 956 were OOV.
  - After the graph update, errors fell to 3,792 and corrected OOV words rose to 1,148.
- **Comparison with earlier work:** reported WER gains are 4.6 (this method), 3.7 (El-Desoky et al., 2009), 1.4 (Afify et al., 2005) and 1.2 (Messaoudi et al., 2006). These figures come from different setups, so they are not directly comparable.

## Interesting ideas
- The lexical resource sits outside the main model and grows continuously from crawled text, so new words reach the system without retraining the language model.
- Errors are flagged simply because a token is absent from a large attested-word store, which gives a cheap and transparent detection rule.
- The context window around an error is widened step by step, and the width that leaves the fewest candidates is chosen, treating a small candidate set as a sign of a tight context fit.
- Instead of committing to one fix, the system outputs every candidate with its left and right succession probabilities, leaving the final choice to later components.
- The corrector is evaluated twice, before and after a resource update, and the gain specifically in fixed OOV words is tracked.

## For Mova

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

- **I1** (world): Compiled knowledge such as the lexicon or entity lists could be kept as an appendable side store, so new names and terms reach the world model without rebuilding everything.
- **I2** (tokens): Unknown-token flagging against a corpus-attested lexicon could route suspicious tokens (typos, OCR noise) to a repair step before tagging, with the reason recorded.
- **I3** (learn): Self-induced repair rules could use the same principle, preferring the smallest context that still disambiguates and logging that context as the rule's explanation.
- **I4** (tree): Ambiguous tokens could be passed to the parser as a weighted lattice of alternatives, letting syntactic agreement pick the reading and keeping the choice explainable.
- **I5** (learn): When the knowledge base or lexicon is extended, the same test could be rerun on a fixed set, with the unseen-item subset reported separately, to confirm that the gain comes from the update itself.

*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

- [University of Monastir](https://md.research.elamur.ai/free/en/universities/university-of-monastir/) — University, Monastir, Tunisia · 8 works · 52 citations

📚 Shelves (7): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works · [💭 Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/) · 1,649 works · [💭 Experimental and Cognitive Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/experimental-and-cognitive-psychology/) · 930 works · [🔊 Signal Processing](https://md.research.elamur.ai/free/en/shelves/computer-science/signal-processing/) · 493 works

🧭 Topics (3): [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · 3,894 works · [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 3,066 works · [🎙️ Speech & spoken language](https://md.research.elamur.ai/free/en/topics/speech/) · 519 works

## Cite

Cite as: Mohamed Labidi, Mohsen Maraoui and Mounir Zrigui. 2017. Unsupervised Method for Improving Arabic Speech Recognition Systems. Proceedings of the 31st Pacific Asia Conference on Language, Information and Computation.

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

```bibtex
@inproceedings{labidi2017unsupervised,
  title = {{Unsupervised Method for Improving Arabic Speech Recognition Systems}},
  author = {Mohamed Labidi, Mohsen Maraoui and Mounir Zrigui},
  booktitle = {Proceedings of the 31st Pacific Asia Conference on Language, Information and Computation},
  year = {2017},
  url = {https://aclanthology.org/Y17-1024/}
}
```

- [← previous Universal Joint Morph-Syntactic Processing: The Open University of Israel's Submission to The CoNLL 2017 Shared Task](https://md.research.elamur.ai/free/en/papers/more-2017-universal-joint-morph-syntactic-processing/)
- [next → Up2B2: Playing English grammar games at the B2 level](https://md.research.elamur.ai/free/en/papers/lvalue-2018-up2b2-playing-english-grammar-games/)
- [↑ back to papers of 2017](https://md.research.elamur.ai/free/en/papers/2017/page/6/)

## References

15 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Mourad Abbas, Kamel Smaïli, and Daoud Berkani. 2011. Evaluation of Topic Identification Methods on Arabic Corpora. JDIM, 9(5):185–192. [find](https://scholar.google.com/scholar?q=Mourad+Abbas%2C+Kamel+Sma%C3%AFli%2C+and+Daoud+Berkani%2E+2011%2E+Evaluation+of+Topic+Identification+Methods+on+Arabic+Corpora%2E+JDIM%2C+9%285%29%3A185%E2%80%93192%2E)
2. Mohamed Afify, Long Nguyen, Bing Xiang, Sherif Abdou, and John Makhoul. 2005. Recent progress in Arabic broadcast news transcription at BBN. In Interspeech, volume 5, pages 1637–1640. [find](https://scholar.google.com/scholar?q=Mohamed+Afify%2C+Long+Nguyen%2C+Bing+Xiang%2C+Sherif+Abdou%2C+and+John+Makhoul%2E+2005%2E+Recent+progress+in+Arabic+broadcast+news+transcription+at+BBN%2E+In+Interspeech%2C+volume+5%2C+pages+1637%E2%80%931640%2E)
3. Sarah Al-Shareef and Thomas Hain. 2012. CRF-based Diacritisation of Colloquial Arabic for Automatic Speech Recognition. In INTERSPEECH, pages 1824–1827. [find](https://scholar.google.com/scholar?q=Sarah+Al%2DShareef+and+Thomas+Hain%2E+2012%2E+CRF%2Dbased+Diacritisation+of+Colloquial+Arabic+for+Automatic+Speech+Recognition%2E+In+INTERSPEECH%2C+pages+1824%E2%80%931827%2E)
4. Mathias Creutz, Teemu Hirsimäki, Mikko Kurimo, Antti Puurula, Janne Pylkkönen, Vesa Siivola, Matti Varjokallio, Ebru Arisoy, Murat Saraçlar, and Andreas Stolcke. 2007. Morph-based speech recognition and modeling of out-of-vocabulary words across languages. ACM Transactions on Speech and Language Processing (TSLP), 5(1):3. [find](https://scholar.google.com/scholar?q=Mathias+Creutz%2C+Teemu+Hirsim%C3%A4ki%2C+Mikko+Kurimo%2C+Antti+Puurula%2C+Janne+Pylkk%C3%B6nen%2C+Vesa+Siivola%2C+Matti+Varjokallio%2C+Ebru+Arisoy%2C+Murat+Sara%C3%A7lar%2C+and+Andreas+Stolcke%2E+2007%2E+Morph%2Dbased+speech+recognition+a)
5. Frank Diehl, Mark JF Gales, Marcus Tomalin, and Philip C Woodland. 2009. Morphological analysis and decomposition for Arabic speech-to-text systems. In INTERSPEECH, pages 2675–2678. [find](https://scholar.google.com/scholar?q=Frank+Diehl%2C+Mark+JF+Gales%2C+Marcus+Tomalin%2C+and+Philip+C+Woodland%2E+2009%2E+Morphological+analysis+and+decomposition+for+Arabic+speech%2Dto%2Dtext+systems%2E+In+INTERSPEECH%2C+pages+2675%E2%80%932678%2E)
6. Amr El-Desoky, Christian Gollan, David Rybach, Ralf Schlüter, and Hermann Ney. 2009. Investigating the use of morphological decomposition and diacritization for improving Arabic LVCSR. In Interspeech, pages 2679–2682. [find](https://scholar.google.com/scholar?q=Amr+El%2DDesoky%2C+Christian+Gollan%2C+David+Rybach%2C+Ralf+Schl%C3%BCter%2C+and+Hermann+Ney%2E+2009%2E+Investigating+the+use+of+morphological+decomposition+and+diacritization+for+improving+Arabic+LVCSR%2E+In+Interspeech%2C)
7. Nizar Habash. 2009. REMOOV: A tool for online handling of out-of-vocabulary words in machine translation. In Proceedings of the 2nd International Conference on Arabic Language Resources and Tools (MEDAR), Cairo, Egypt. [find](https://scholar.google.com/scholar?q=Nizar+Habash%2E+2009%2E+REMOOV%3A+A+tool+for+online+handling+of+out%2Dof%2Dvocabulary+words+in+machine+translation%2E+In+Proceedings+of+the+2nd+International+Conference+on+Arabic+Language+Resources+and+Tools+%28MED)
8. Paul Lamere, Philip Kwok, William Walker, Evandro B Gouvêa, Rita Singh, Bhiksha Raj, and Peter Wolf. 2003. Design of the CMU sphinx-4 decoder. In INTERSPEECH. Citeseer. [find](https://scholar.google.com/scholar?q=Paul+Lamere%2C+Philip+Kwok%2C+William+Walker%2C+Evandro+B+Gouv%C3%AAa%2C+Rita+Singh%2C+Bhiksha+Raj%2C+and+Peter+Wolf%2E+2003%2E+Design+of+the+CMU+sphinx%2D4+decoder%2E+In+INTERSPEECH%2E+Citeseer%2E)
9. In this paper we have tried to deal with the challenges of the limit of vocabulary and the Beth Logan, J-M Van Thong, and Pedro J Moreno. 2005. Approaches to reduce the effects of OOV Arabic language characteristics in large queries on indexed spoken audio. IEEE vocabulary Arabic speech recognition systems. transactions on multimedia, 7(5):899–906. We have tested a graph-based method. It has given a good reduction by 4.6% in terms of Abdelkhalek Messaoudi, J Gauvain, and Lori Lamel. 2006. Arabic broadcast news transcription using a WER. Furthermore, it has fairly dealt with the one million word vocalized vocabulary. In Arabic language characteristics. The proposed 167 Acoustics, Speech and Signal Processing, 2006. method presents a good step in this field and in ICASSP 2006 Proceedings. 2006 IEEE International Conference on, volume 1, pages I–I. IEEE. [find](https://scholar.google.com/scholar?q=In+this+paper+we+have+tried+to+deal+with+the+challenges+of+the+limit+of+vocabulary+and+the+Beth+Logan%2C+J%2DM+Van+Thong%2C+and+Pedro+J+Moreno%2E+2005%2E+Approaches+to+reduce+the+effects+of+OOV+Arabic+language+)
10. Kenney Ng and Victor W Zue. 2000. Subword-based approaches for spoken document retrieval. Speech Communication, 32(3):157–186. [find](https://scholar.google.com/scholar?q=Kenney+Ng+and+Victor+W+Zue%2E+2000%2E+Subword%2Dbased+approaches+for+spoken+document+retrieval%2E+Speech+Communication%2C+32%283%29%3A157%E2%80%93186%2E)
11. Scott Novotney, Richard M Schwartz, and Sanjeev Khudanpur. 2011. Unsupervised Arabic Dialect Adaptation with Self-Training. In INTERSPEECH, pages 541–544. [find](https://scholar.google.com/scholar?q=Scott+Novotney%2C+Richard+M+Schwartz%2C+and+Sanjeev+Khudanpur%2E+2011%2E+Unsupervised+Arabic+Dialect+Adaptation+with+Self%2DTraining%2E+In+INTERSPEECH%2C+pages+541%E2%80%93544%2E)
12. Majid Razmara, Maryam Siahbani, Reza Haffari, and Anoop Sarkar. 2013. Graph Propagation for Paraphrasing Out-of-Vocabulary Words in Statistical Machine Translation. In ACL (1), pages 1105–1115. Citeseer. [find](https://scholar.google.com/scholar?q=Majid+Razmara%2C+Maryam+Siahbani%2C+Reza+Haffari%2C+and+Anoop+Sarkar%2E+2013%2E+Graph+Propagation+for+Paraphrasing+Out%2Dof%2DVocabulary+Words+in+Statistical+Machine+Translation%2E+In+ACL+%281%29%2C+pages+1105%E2%80%931115%2E+Citese)
13. Motaz K Saad and Wesam Ashour. 2010. Arabic morphological tools for text mining. Corpora, 18:19. [find](https://scholar.google.com/scholar?q=Motaz+K+Saad+and+Wesam+Ashour%2E+2010%2E+Arabic+morphological+tools+for+text+mining%2E+Corpora%2C+18%3A19%2E)
14. Andreas Stolcke and others. 2002. SRILM-an extensible language modeling toolkit. In Interspeech, volume 2002, page 2002. [find](https://scholar.google.com/scholar?q=Andreas+Stolcke+and+others%2E+2002%2E+SRILM%2Dan+extensible+language+modeling+toolkit%2E+In+Interspeech%2C+volume+2002%2C+page+2002%2E)
15. Taha Zerrouki and Amar Balla. 2017. Tashkeela: Novel corpus of Arabic vocalized texts, data for auto-diacritization systems. Data in Brief, 11:147– 151. [find](https://scholar.google.com/scholar?q=Taha+Zerrouki+and+Amar+Balla%2E+2017%2E+Tashkeela%3A+Novel+corpus+of+Arabic+vocalized+texts%2C+data+for+auto%2Ddiacritization+systems%2E+Data+in+Brief%2C+11%3A147%E2%80%93+151%2E)

*Updated 2 November 2026*