Elamur Research⌕
☰
home / papers / 2009 / Spoken Arabic Dialect Identification Using Phonotactic Modeling

Spoken Arabic Dialect Identification Using Phonotactic Modeling

Paper · 2009 · Fadi Biadsy, Julia Hirschberg and Nizar Habash

Summary

The paper addresses automatic identification of Arabic varieties from speech alone: four colloquial dialects (Gulf, Iraqi, Levantine, Egyptian) plus MSA. Dialect ID is harder than language ID because the varieties are closely related. Knowing the variety would let ASR systems switch acoustic, pronunciation and language models, and could also support speaker profiling and dialect-aware TTS. The authors adapt parallel PRLM, a phonotactic technique from language ID. Several phone recognizers convert audio into phone strings, phone n-gram models are trained per dialect, and a discriminative classifier combines the resulting perplexities into a decision. The dialect data are LDC telephone conversations, used without their transcripts. MSA comes from Broadcast News. The recognizers cover six foreign languages and three MSA variants. With 30-second test segments, five-way accuracy is 81.60%, and it rises with longer input. MSA is almost never confused with the dialects. Iraqi and Gulf are the hardest pair to separate. The study shows that phonotactics alone, without any lexical transcripts, carries substantial dialect signal.

Key points

  • Phone recognizers: English, German, Hindi, Japanese, Mandarin and Spanish, plus three MSA ones:
    • ArbO: open-loop, no emphatic vowels.
    • ArbOE: open-loop, with emphatic vowels.
    • ArbLME: emphatic vowels plus a bigram phone LM.
  • Five-way accuracy grows with test duration: 68.67% at 5s, 81.60% at 30s, 86.93% at 60s and 87.86% at 120s.
  • F-Measure at 30s:
    • MSA: above 98% at every duration.
    • Egyptian: 90.2%.
    • Levantine: 79.4%.
    • Iraqi: 71.7%.
    • Gulf: 68.3%.
    • The Iraqi–Gulf confusion matches classifications that treat Iraqi as a Gulf sub-dialect.
  • The Arabic-trained recognizers matter most, especially those modeling emphatic vowels. Removing all of them causes a significant accuracy drop.
  • The authors flag that the MSA results may be inflated:
    • The MSA data differ in genre (Broadcast News, partly read speech).
    • The MSA test speaker identities were assigned automatically.

Interesting ideas

  • I1. Mismatched tokenizers as feature generators: recognizers trained on unrelated languages still yield phone streams whose sequential statistics separate dialects, so an imperfect front-end can be useful if its errors are consistent.
  • I2. Stacking perplexities instead of argmax: the scores of all per-dialect models from all streams feed a discriminative classifier, rather than simply picking the most likely model. The authors single out this classifier as an important component.
  • I3. Greedy backward pruning of components: drop one recognizer at a time and keep it out only if cross-validated accuracy does not fall. This beat automatic feature selection here, and a subset outperformed using everything.
  • I4. Reporting accuracy as a function of input length exposes how much evidence the method needs, which is more informative than a single number.
  • I5. Self-flagged confounds: the near-perfect MSA score is attributed partly to genre and recording differences rather than the target signal.

For Mova

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

  • I1 (tree): Parse a text with several parsers trained on different treebanks and use each one's per-sentence model scores as a signature for detecting English variety or genre before choosing the UD dialect conversion.
  • I2 (learn): Feed the confidence scores of all witness parsers into a small learned arbiter, rather than relying on majority vote, when agreeing on silver trees.
  • I3 (learn): Apply the same leave-one-out loop to witness parsers or self-induced repair rules, keeping only those whose removal hurts held-out accuracy.
  • I4 (answer): Report reader accuracy on fables and tests bucketed by text length, to show where world-building from short contexts breaks down.
  • I5 (learn): Add negative controls that keep the label fixed but vary the source or genre, so a classifier or evaluator cannot win by detecting the corpus instead of the phenomenon.

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

🎓Columbia UniversityNew York · United States241 works · 54,053 citations

 Fields, topics and concepts

📚 Shelves 6The 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 works

🧭 Topics 1Research subjects and languages the work is about.Semitic languages · 2,908 works

Cite

Cite as Fadi Biadsy, Julia Hirschberg and Nizar Habash. 2009. Spoken Arabic Dialect Identification Using Phonotactic Modeling. Proceedings of the EACL 2009 Workshop on Computational Approaches to Semitic Languages.

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

@inproceedings{biadsy2009spoken,
  title = {{Spoken Arabic Dialect Identification Using Phonotactic Modeling}},
  author = {Fadi Biadsy, Julia Hirschberg and Nizar Habash},
  booktitle = {Proceedings of the EACL 2009 Workshop on Computational Approaches to Semitic Languages},
  year = {2009},
  doi = {10.3115/1621774.1621784},
  url = {https://doi.org/10.3115/1621774.1621784}
}

References

22 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. F. S. Alorfi. 2008. PhD Dissertation: Automatic Identification Of Arabic Dialects Using Hidden Markov Models. In University of Pittsburgh. find
  2. Appen Pty Ltd. 2006a. Gulf Arabic Conversational Telephone Speech Linguistic Data Consortium, Philadelphia. find
  3. Appen Pty Ltd. 2006b. Iraqi Arabic Conversational Telephone Speech Linguistic Data Consortium, Philadelphia. find
  4. M. Barkat, J. Ohala, and F. Pellegrino. 1999. Prosody as a Distinctive Feature for the Discrimination of Arabic Dialects. In Proceedings of Eurospeech’99. find
  5. F. Biadsy, N. Habash, and J. Hirschberg. 2009. Improving the Arabic Pronunciation Dictionary for Phone and Word Recognition with Linguistically-Based Pronunciation Rules. In Proceedings of NAACL/HLT 2009, Colorado, USA.
  6. A. Canavan and G. Zipperlen. 1996. CALLFRIEND Egyptian Arabic Speech Linguistic Data Consortium, Philadelphia. find
  7. A. Canavan, G. Zipperlen, and D. Graff. 1997. CALL-HOME Egyptian Arabic Speech Linguistic Data Consortium, Philadelphia. find
  8. J. S. Garofolo et al. 1993. TIMIT Acoustic-Phonetic Continuous Speech Corpus Linguistic Data Consortium, Philadelphia. find
  9. N. Habash. 2006. On Arabic and its Dialects. Multilingual Magazine, 17(81). find
  10. R. Hamdi, M. Barkat-Defradas, E. Ferragne, and F. Pellegrino. 2004. Speech Timing and Rhythmic Structure in Arabic Dialects: A Comparison of Two Approaches. In Proceedings of Interspeech’04. find
  11. C. Holes. 2004. Modern Arabic: Structures, Functions, and Varieties. Georgetown University Press. Revised Edition. find
  12. B. Ma, D. Zhu, and R. Tong. 2006. Chinese Dialect Identification Using Tone Features Based On Pitch Flux. In Proceedings of ICASP’06. find
  13. M. Maamouri. 2006. Levantine Arabic QT Training Data Set 5, Speech Linguistic Data Consortium, Philadelphia. find
  14. P. Matejka, P. Schwarz, J. Cernocky, and P. Chytil. 2005. Phonotactic Language Identification using High Quality Phoneme Recognition. In Proceedings of Eurospeech’05. find
  15. Y. K. Muthusamy, R.A. Cole, and B.T. Oshika. 1992. The OGI Multi-Language Telephone Speech Corpus. In Proceedings of ICSLP’92. find
  16. J. Peters, P. Gilles, P. Auer, and M. Selting. 2002. Identification of Regional Varieties by Intonational Cues. An Experimental Study on Hamburg and Berlin German. 45(2):115–139. ↗
  17. F. Ramus. 2002. Acoustic Correlates of Linguistic Rhythm: Perspectives. In Speech Prosody. find
  18. A. Stolcke. 2002. SRILM - an Extensible Language Modeling Toolkit. In ICASP’02, pages 901–904. find
  19. P. Torres-Carrasquillo, T. P. Gleason, and D. A. Reynolds. 2004. Dialect identification using Gaussian Mixture Models. In Proceedings of the Speaker and Language Recognition Workshop, Spain. find
  20. S. Young, G. Evermann, M. Gales, D. Kershaw, G. Moore, J. Odell, D. Ollason, D. Povey, V. Valtchev, and P. Woodland. 2006. The HTK Book, version 3.4. find
  21. M. A. Zissman, T. Gleason, D. Rekart, and B. Losiewicz. 1996. Automatic Dialect Identification of Extemporaneous Conversational, Latin American Spanish Speech. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing, Atlanta, USA. find
  22. M. A. Zissman. 1996. Comparison of Four Approaches to Automatic Language Identification of Telephone Speech. IEEE Transactions of Speech and Audio Processing, 4(1). ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/biadsy-2009-spoken-arabic-dialect-identification-phonotactic/