Elamur Research⌕
☰
home / papers / 2017 / An enhanced automatic speech recognition system for Arabic

An enhanced automatic speech recognition system for Arabic

Paper · 2017 · Mohamed Amine Menacer, Odile Mella, Dominique Fohr et al.

Summary

Arabic speech recognition is hard for several reasons. The morphology is rich, so many words fall outside the vocabulary. Written text usually omits short vowels, so one written form can have many pronunciations. The hamoza sign is often confused or left out. The authors build a Modern Standard Arabic (MSA) recognizer with the Kaldi toolkit, trained on about 63 hours of broadcast news from the ELRA corpora Nemlar and NetDC. The language models are n-grams estimated on Arabic GigaWord plus the training transcripts, after rule-based text normalization. They compare GMM-HMM and DNN-HMM acoustic models with adaptation and discriminative training (fMLLR, SAT, MMI, sMBR).

Two post-processing steps follow:

  • rescoring the recognition lattice with a larger 4-gram LM;
  • fixing hamoza placement above or below Alif, using dictionary lookup plus word2vec context similarity.

The baseline reaches a WER of 14.42 on Test; combining both steps brings it to 13.45. The authors report this as about 12% relative gain, although the table figures imply a smaller one. The work shows that a cheap, targeted correction of one dominant error type adds real value on top of a standard pipeline.

Key points

  • Acoustic data: Nemlar and NetDC (ELRA), 16 kHz news broadcasts, 63 hours in total, split 52 / 6 / 5 hours into Train / Dev / Test. Pronunciations use Arabic SAMPA lexicons with 34 phonemes.
  • Language model data: Arabic GigaWord (about 1,000 million words) plus roughly 315k words of acoustic transcripts. Short vowels are stripped, and numbers, times and abbreviations are spelled out as words. Farasa splits off the "wa" prefix.
  • Best acoustic model: a DNN with 6 hidden layers of 2048 units, 440 inputs (spliced fMLLR features) and 4,264 output states, about 30.6M parameters. It is trained first with cross-entropy, then with sMBR.
  • Lattice rescoring: rescoring with the full 4-gram LM gained 1.58 (Dev) and 1.3 (Test) absolute over lattices produced with a heavily pruned 4-gram. A pruned 2-gram with more n-grams beat that aggressively pruned 4-gram.
  • Hamoza correction:
    • Detection: a 9.2-million-entry word list from Attia et al. (2012).
    • Candidates: words that differ only in the hamoza form on Alif.
    • Selection: cosine similarity over 200-dimensional cbow vectors trained on GigaWord.
    • Effect on Test WER: correction alone 14.14, rescoring alone 14.02, both together 13.45.

Interesting ideas

  • I1. Analyse the errors on the dev set first, find the single most frequent error class, and then build a narrow post-fix just for it. This beats generic tuning when one error type dominates.
  • I2. Generate correction candidates from a closed confusion set (here, only the hamoza variants) instead of a general edit distance. This keeps the candidate list tiny and the decision easy to explain.
  • I3. Use a two-pass design: decode cheaply with a small model, then rescore the full lattice of hypotheses with a heavier model.
  • I4. The finding that a less-pruned 2-gram beats a harshly pruned 4-gram shows that coverage can matter more than model order. Before scaling up a model, check what pruning removed.
  • I5. Normalize text into spoken words before modelling: numbers, times and abbreviations become their literal word forms, and noise and stretched spellings are removed. This shrinks sparse surface variation.

For Mova

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

  • I1 (learn): Mova's self-induced repair rules could be ranked by how many dev errors each class explains, so effort goes to the largest error bucket first.
  • I2 (tokens): lemma and spelling repair could use explicit confusion sets (its/it's, then/than, homographs), giving an auditable small choice instead of fuzzy matching.
  • I3 (tree): the perceptron parser could emit k-best or packed alternatives, which gparse or a frame-consistency check then rescores, keeping each pass inspectable.
  • I4 (world): when compiling verb classes or knowledge-graph slices, measure the coverage lost to pruning before choosing a richer but thinner representation.
  • I5 (answer): maths word problems and test readers could canonicalize numerals, times and units into one normalized form, so quantity frames are built consistently.

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

🔬Laboratoire Lorrain de Recherche en Informatique et ses ApplicationsNancy · France30 works · 3,899 citations

 Fields, topics and concepts

📚 Shelves 7The 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 worksPsychology · 1,552 worksExperimental and Cognitive Psychology · 870 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 Mohamed Amine Menacer, Odile Mella, Dominique Fohr, Denis Jouvet, David Langlois and Kamel Smaili. 2017. An enhanced automatic speech recognition system for Arabic. Proceedings of the Third Arabic Natural Language Processing Workshop.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/w17-1319

@inproceedings{menacer2017enhanced,
  title = {{An enhanced automatic speech recognition system for Arabic}},
  author = {Mohamed Amine Menacer, Odile Mella, Dominique Fohr, Denis Jouvet, David Langlois and Kamel Smaili},
  booktitle = {Proceedings of the Third Arabic Natural Language Processing Workshop},
  year = {2017},
  doi = {10.18653/v1/w17-1319},
  url = {https://doi.org/10.18653/v1/w17-1319}
}

References

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

  1. Jeff A. Bilmes and Katrin Kirchhoff. 2003. Factored language models and generalized parallel backoff. In Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology: companion volume of the Proceedings of HLT-NAACL 2003–short papers-Volume 2, pages 4–6. Association for Computational Linguistics. find
  2. Ahmed Abdelali, Kareem Darwish, Nadir Durrani, and Hamdy Mubarak. 2016. Farasa: A fast and furious segmenter for arabic. In HLT-NAACL Demos, pages 11–16. Association for Computational Linguistics, San Diego, California.
  3. Ciprian Chelba, Thorsten Brants, Will Neveitt, and Peng Xu. 2010. Study on interaction between entropy pruning and kneser-ney smoothing. pages 2422–2425. find
  4. Stanley F. Chen and Joshua Goodman. 1996. An empirical study of smoothing techniques for language modeling. In Proceedings of the 34th annual meeting on Association for Computational Linguistics, pages 310–318. Association for Computational Linguistics. ↗
  5. Mohamed Afify, Ruhi Sarikaya, Hong-Kwang Jeff Kuo, Laurent Besacier, and Yuqing Gao. 2006. On the use of morphological analysis for dialectal arabic speech recognition. In Interspeech. find
  6. 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 Spoken Language Technology Workshop (SLT), 2014 IEEE, pages 525–529, Dec. find
  7. Ghinwa Choueiter, Daniel Povey, Stanley F. Chen, and Geoffrey Zweig. 2006. Morpheme-based language modeling for arabic lvcsr. In 2006 IEEE International Conference on Acoustics Speech and Signal Processing Proceedings, volume 1, pages I–I. IEEE. find
  8. Mohamed I. Alkanhal, Mohamed A. Al-Badrashiny, Mansour M. Alghamdi, and Abdulaziz O. Al-Qabbany. 2012. Automatic stochastic arabic spelling correction with emphasis on space insertions and deletions. IEEE Transactions on Audio, Speech, and Language Processing, 20(7):2111– 2122. ↗
  9. Frank Diehl, Mark J. F. Gales, Marcus Tomalin, and Philip C. Woodland. 2009. Morphological analysis and decomposition for arabic speech-to-text systems. In Interspeech, pages 2675–2678. find
  10. 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
  11. Nouf AlShenaifi, Rehab AlNefie, Maha Al-Yahya, and Hend Al-Khalifa. 2015. Arib@ qalb-2015 shared task: A hybrid cascade model for arabic spelling error detection and correction. In ANLP Workshop 2015, page 127.
  12. Khaled Shaalan, Rana Aref, and Aly Fahmy. 2010. An approach for analyzing and correcting spelling errors for non-native arabic learners. In Informatics and Systems (INFOS), 2010 The 7th International Conference on, pages 1–7. IEEE. find
  13. Amr El-Desoky, Ralf Schlüter, and Hermann Ney. 2010. A hybrid morphologically decomposed factored language models for arabic lvcsr. In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 701–704. Association for Computational Linguistics.
  14. Andreas Stolcke. 2000. Entropy-based pruning of backoff language models. arXiv preprint cs/0006025. find
  15. Youssef Hassan, Mohamed Aly, and Amir Atiya. 2014. Arabic spelling correction using supervised learning. arXiv preprint arXiv:1409.8309. ↗
  16. Andreas Stolcke. 2002. SRILM-an extensible language modeling toolkit. In Interspeech, volume 2002, page 2002. find
  17. Slava Katz. 1987. Estimation of probabilities from sparse data for the language model component of a speech recognizer. IEEE transactions on acoustics, speech, and signal processing, 35(3):400–401. ↗
  18. Bing Xiang, Kham Nguyen, Long Nguyen, Richard Schwartz, and John Makhoul. 2006. Morphological decomposition for arabic broadcast news transcription. In 2006 IEEE International Conference on Acoustics Speech and Signal Processing Proceedings, volume 1, pages I–I. IEEE. find
  19. Katrin Kirchhoff et al. 2002. Novel speech recognition models for arabic. In Johns-Hopkins University summer research workshop. find
  20. Taha Zerrouki, Khaled Alhawaity, and Amar Balla. 2014. Autocorrection of arabic common errors for large text corpus. ANLP 2014, page 127. find
  21. Lori Lamel, Abdelkhalek Messaoudi, and Jean-Luc Gauvain. 2009. Automatic speech-to-text transcription in arabic. ACM Transactions on Asian Language Information Processing (TALIP), 8(4):18. ↗
  22. Young-Suk Lee, Kishore Papineni, Salim Roukos, Ossama Emam, and Hany Hassan. 2003. Language model based arabic word segmentation. In Proceedings of the 41st Annual Meeting on Association for Computational Linguistics-Volume 1, pages 399–406. Association for Computational Linguistics.
  23. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Distributed representations of words and phrases and their compositionality. arXiv preprint arXiv:1310.4546. ↗
  24. Mehryar Mohri, Fernando Pereira, and Michael Riley. 2008. Speech recognition with weighted finite-state transducers. In Springer Handbook of Speech Processing, pages 559–584. Springer. ↗
  25. Tim Ng, Kham Nguyen, Rabih Zbib, and Long Nguyen. 2009. Improved morphological decomposition for arabic broadcast news transcription. In 2009 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4309–4312. IEEE. find
  26. Daniel Povey and George Saon. 2006. Feature and model space speaker adaptation with full covariance gaussians. In Interspeech. find
  27. Daniel Povey, Arnab Ghoshal, Gilles Boulianne, Lukas Burget, Ondrej Glembek, Nagendra Goel, Mirko Hannemann, Petr Motlicek, Yanmin Qian, Petr Schwarz, Jan Silovsky, Georg Stemmer, and Karel Vesely. 2011. The kaldi speech recognition toolkit. In IEEE 2011 Workshop on Automatic Speech Recognition and Understanding. IEEE Signal Processing Society, December. IEEE Catalog No.: CFP11SRW-USB. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/menacer-2017-enhanced-automatic-speech-recognition-system/