Building an automatic speech recognition system for Arabic based on CMU Sphinx

Original (French): Realisation d'un systeme de reconnaissance automatique de la parole arabe base sur CMU Sphinx

Paper · 2010 · Ali Sadiqui and Noureddine Chenfour

Summary

This paper, written in French, describes building an automatic speech recognition (ASR) system for Arabic using the open-source CMU Sphinx toolkit. The authors address the limited ASR tooling for Arabic by adapting an existing, general-purpose recognizer instead of building one from scratch. The work appears to cover the standard Sphinx pipeline: preparing Arabic speech data, training models, decoding, and reporting recognition results. Page 11 contains a large results table. The extracted PDF text is heavily font-corrupted, so the specific corpus, model configuration and numeric scores cannot be reliably recovered and are not reported here. The paper's main value is as a case study of porting a mature, language-independent ASR framework to an under-served language. That is a practical route for any team bootstrapping speech technology for a new language.

Key points

Authors: Ali Sadiqui and Noureddine Chenfour. It appeared in the Annals of the University of Tibiscus, Computer Science Series, volume VIII, issue 1 (2010), on pages 27–40, and is also on arXiv as 1202.6266.
The system is built on CMU Sphinx, an open-source, statistically trained speech recognition framework.
The target language is Arabic, which the authors treat as under-resourced for ASR.
Results are given as a per-item table (p. 11). The figures are unreadable in the extracted text, so no accuracy numbers are cited here.
The paper is written in French, which places it in the North-African francophone research context.

Interesting ideas

I1. Port an existing language-independent toolkit to a new language and invest the effort in language-specific resources, not in a new engine
I2. Report results per item in a detailed table, not only as one aggregate score, so that systematically hard cases stand out
I3. Treat a mature open framework's documented pipeline (data preparation, training, decoding) as a reproducible recipe for a low-resource setting

For Mova

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

I1 (tokens): the same strategy could extend crate en's tokenizer and morphology to new varieties by swapping lexicons and rules while keeping the core machinery untouched.
I2 (learn): evaluation runs could publish per-construction or per-label breakdowns, which makes regressions and weak spots explainable rather than hidden in an average.
I3 (tree): a fixed, scripted recipe for training gparse or the perceptron parser on a new treebank would make UD dialect experiments repeatable and comparable.

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

📚 Shelves 5Computer Science · 9,086 works Artificial Intelligence · 8,889 works Arts and Humanities · 9,296 works Language and Linguistics · 9,206 works Signal Processing · 213 works

🧭 Topics 2Semitic languages · 1,484 works Speech & spoken language · 233 works

Cite

Cite as Ali Sadiqui and Noureddine Chenfour. 2010. Realisation d'un systeme de reconnaissance automatique de la parole arabe base sur CMU Sphinx. arXiv:1202.6266 · Ann. Univ. Tibiscus Comp. Sci. Series VIII / 1 (2010), 27-40.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.1202.6266 · arXiv 1202.6266

@misc{sadiqui2010realisation,
  title = {{Realisation d'un systeme de reconnaissance automatique de la parole arabe base sur CMU Sphinx}},
  author = {Ali Sadiqui and Noureddine Chenfour},
  year = {2010},
  doi = {10.48550/arxiv.1202.6266},
  eprint = {1202.6266},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1202.6266}
}

Updated 24 October 2026

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sadiqui-2012-realisation-d-un-systeme-de/