Elamur Research⌕
☰
home / papers / 2007 / Arabic Speech Recognition System using CMU-Sphinx4

Arabic Speech Recognition System using CMU-Sphinx4

Paper · 2007 · H. Satori, M. Harti and N. Chenfour

Summary

The authors build one of the first Arabic speech recognizers on top of the open-source CMU Sphinx-4 toolkit, a Java framework built on hidden Markov models. Despite the English title, the paper is written in French. It is mainly a proof of concept. The recognition engine stays as it is, and only the two language-dependent parts are rebuilt: the acoustic model and the language model.

To do this, the authors:

  • defined a phoneme symbol set for Classical Arabic,
  • wrote a pronunciation dictionary for the digits 0–9,
  • trained HMM acoustic models on MFCC features with SphinxTrain,
  • restricted recognition to digit strings with a simple JSGF grammar.

The training corpus is small: six Moroccan speakers each said every digit five times. Test speakers of both sexes reached about 80–87% recognition accuracy, with a slightly higher average for men.

The work matters as a recipe. It shows that a modest research group can port a mature open ASR toolkit to an under-served language by swapping in language-specific resources. Planned next steps are large-vocabulary Arabic and Moroccan dialect recognition.

Key points

  • Toolkit: CMU Sphinx-4 (Java, HMM-based) handles decoding and SphinxTrain handles acoustic-model training. The system is configured through an XML file.
  • Arabic phonology: the paper describes Classical Arabic as having 34 phonemes (6 vowels and 28 consonants), with pharyngeal and emphatic classes. Allowed syllables are CV, CVC and CVCC.
  • Corpus "Arabic digits":
    • content: the 10 digits, 6 Moroccan speakers (3 male, 3 female), 5 repetitions each, 300 tokens in total;
    • format: 16 kHz, 16-bit, mono WAV;
    • quality check: every take was played back to confirm the whole digit was captured.
  • Models: the acoustic model is trained on MFCC feature vectors using a custom phoneme symbol table and a word-to-phoneme dictionary. The language model is a JSGF grammar that accepts any sequence of the ten digits.
  • Results:
    • each speaker read the ten digits in three trials; per-speaker accuracy ranged from 80.00% to 86.66%;
    • the male average was 85.56% and the female average was 83.34%;
    • the authors attribute the gap to perfect accuracy to the tiny training set, citing a recommendation of 500+ voices;
    • the paper does not say clearly whether the test speakers overlap with the training speakers.

Interesting ideas

  • I1. Swap only the language-dependent parts. Treat the engine as fixed and port it to a new language by replacing just the language-specific resources (here the acoustic model and the grammar), so the adaptation stays small and auditable.
  • I2. Use an explicit hand-written lexicon as the interface. A word-to-phoneme dictionary links each output word to its units in a form a human can read and check, which supports explainability.
  • I3. Constrain the output space with a task grammar. A tiny grammar listing only the legal outputs shrinks the search space and removes impossible answers before scoring.
  • I4. Break results down by subgroup. Reporting accuracy per speaker and per gender, rather than as one figure, exposes uneven performance across groups.
  • I5. Run a cheap completeness check during data collection. Replaying each recording to confirm the full item was captured is a low-cost quality gate on the raw data.

For Mova

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

  • I1 (tree): The parser core could stay fixed while UD dialect or treebank-specific conventions live in swappable, declaratively configured conversion tables.
  • I2 (tokens): Keep lemma and morphology lexicon entries as inspectable records, so every tagging or lemma decision can be traced to a specific entry.
  • I3 (answer): Maths word-problem and test readers could restrict candidate answers to a typed grammar (a number with a unit, or one of the listed options) before choosing.
  • I4 (learn): Evaluations could report scores per genre, source or text type, so a strong overall average cannot hide a weak slice.
  • I5 (learn): Silver data could be spot-checked for truncated sentences or missing tokens before training, using a simple automatic completeness test.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 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 H. Satori, M. Harti and N. Chenfour. 2007. Arabic Speech Recognition System using CMU-Sphinx4. arXiv:0704.2201.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.0704.2201 · arXiv 0704.2201

@misc{satori2007arabic,
  title = {{Arabic Speech Recognition System using CMU-Sphinx4}},
  author = {H. Satori, M. Harti and N. Chenfour},
  year = {2007},
  doi = {10.48550/arxiv.0704.2201},
  eprint = {0704.2201},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.0704.2201}
}

References

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

  1. C. Barras, “Reconnassance de la parole continue : adaptation au locuteur et contrôle temporel dans les modèles de markov cachés,” Ph.D. Thesis, Paris VI University, 1996. find
  2. S. Young, “The HTK hidden Markov model toolkit: Design and philosophy,” Cambridge University Engineering Department, UK, Tech. Rep. CUED/FINFENG/TR152, Sept. 1994. find
  3. N. Deshmukh, A. Ganapathiraju, J. Hamaker, J. Picone, and M. Ordowski, “A public domain speech-to-text system,” in Proc. 6th European Conf. Speech Communication and Technology, vol. 5, Budapest, Hungary, Sept. 1999, pp. 2127–2130. find
  4. X. X. Li, Y. Zhao, X. Pi, L. H. Liang, and A. V. Nefian, “Audio-visual continuous speech recognition using a coupled hidden Markov model,” in Proc. 7th International Conf. Spoken Language Processing, Denver, CO, Sept. 2002, pp. 213–216. find
  5. K. F. Lee, H. W. Hon, and R. Reddy, “An overview of the SPHINX speech recognition system,” IEEE Transactions on Acoustics, Speech and Signal Processing, vol. 38, no. 1, pp. 35–45, Jan. 1990. ↗
  6. X. Huang, F. Alleva, H. W. Hon, M. Y. Hwang, and R. Rosenfeld, “The SPHINX-II speech recognition system: an overview,” Computer Speech and Language, vol. 7, no. 2, pp. 137–148, 1993. ↗
  7. M. K. Ravishankar, “Efficient algorithms for speech recognition,” PhD Thesis (CMU Technical Report CS-96-143), Carnegie Mellon University, Pittsburgh, PA, 1996. find
  8. X.D. Huang, Y. Ariki, M.A. Jack, “Hidden Markov models for speech recognition,” Edinburgh: Edinburgh University Press, C, 1990. find
  9. H. Satori M. Harti and N. Chenfour, "Introduction to Arabic Speech Recognition Using CMU SphinxSystem" submitted to int. jour. of comp. sc. Appl. (2007). find
  10. H. Satori M. Harti and N. Chenfour, "Arabic Speech Recognition System based on CMUSphinx" ISCIII 2007 Agadir Maroc. (28-30 Mars 2007). find
  11. G. Z. Hong “Speech Recognition Techniques for Digital Video Library,” University of Hong Kong , 2002. find
  12. X.D. Huang, “The SPHINX-II Speech Recognition System: An Overview,” Computer Speech and Language, Vol. 2, 1993; K. F. Lee, “Automatic Speech Recognition the Development of the SPHINX System,” Kluwer Academic Publishers, 1989. find
  13. M. Al-Zabibi, “An Acoustic–Phonetic Approach in Automatic Arabic Speech Recognition,” The British Library in Association with UMI, 1990. ↗
  14. A. Muhammad, “Alaswaat Alaghawaiyah,” Daar Alfalah, Jordan, 1990 (in Arabic). find
  15. J. Deller, J. Proakis, J.H. Hansen, “Discrete-Time Processing of Speech Signal,” Macmillan, NY, 1993. find
  16. M. Elshafei, “Toward an arabic text-to-speech system,” The Arabian J. Science and Engineering vol. 4B no. 16, pp. 565–583, 1991. find
  17. Y.A. El-Imam, “An unrestricted vocabulary arabic speech synthesis system”, IEEE Transactions on Acoustic, Speech, and Signal Processing vol. 37 , no. 12, pp.1829–1845, 1989. ↗
  18. Y. Ajami Alotaibi “Investigating spoken Arabic digits in speech recognition setting” Inf. and Comp. Sc. pp.173, 115, 2005. ↗
  19. S. Baloul, “Développement d’un système automatique de synthèse de la parole à partir du texte arabe standard voyellé,” Thèse de Doctorat, Université de Maine, Le Mans, 2003. find
  20. A. Varela, H. Cuayáhuitl and J.A. Nolazco-Flores “Creating a Mexican Spanish Version of the CMU Sphinx-III Speech Recognition System” Springer , Vol. 2905, 2003. find
  21. X. Huang, A. Acero, H. Hon, "Spoken language processing a guide to theory, algorithm and system design",Prentice Hall 2001. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/satori-2007-arabic-speech-recognition-system-using/