Elamur Research⌕
☰
home / papers / 2015 / Automatic Dialect Detection in Arabic Broadcast Speech

Automatic Dialect Detection in Arabic Broadcast Speech

Paper · 2015 · Ahmed Ali, Najim Dehak, Patrick Cardinal et al.

Summary

The paper tackles automatic identification of Arabic dialects from broadcast speech. It treats this as a harder variant of language identification, since the varieties belong to one language yet differ in lexicon and sound. The authors compare three utterance representations:

  • sparse count vectors of phone n-grams ("senones", n ≤ 4) taken from a phoneme recognizer;
  • word and word-trigram count vectors taken from an ASR system;
  • i-vectors built on bottleneck (BN) features from a two-stage DNN.

They test generative and discriminative classifiers and fuse the feature types with a multi-class SVM. Training data is broadcast news in EGY, GLF, LAV, NOR and MSA. The test set comes from a different recording channel (Al Jazeera, via QATS) and was labeled through crowdsourcing. Both binary tasks, Arabic vs English and MSA vs dialect, reach 100% accuracy. The five-way task reaches about 59–60%: the abstract reports 59.2% and the conclusion reports 60.2% for the combined system. Many errors trace back to speakers switching between dialect and MSA. The data and the system are released publicly, which gives the field a reproducible benchmark.

Key points

  • Five target classes: Egyptian (EGY), Gulf (GLF), Levantine (LAV), North African (NOR) and MSA, plus English for the language-ID sanity check.
  • On lexical vectors (55K vocabulary, no dimensionality reduction), a linear SVM did best with 45.2% accuracy. The n-gram language model reached 40.4%, Maximum Entropy about 40% and Naive Bayes 37.9%.
  • Feature setup for each representation:
    • Senone vectors use 36 phonemes.
    • The word-trigram dictionary holds 580k entries.
    • The i-vector system uses a 2048-component UBM, 400-dimensional i-vectors and LDA plus WCCN.
    • The BN extractor was trained on 60 hours of Al-Jazeera MSA.
  • Test labels were collected on CrowdFlower. Each file needed at least 3 judges, with collection stopping at 9 judges or at 75% agreement.
  • Code and data are released at github.com/Qatar-Computing-Research-Institute/dialectID and on the QCRI portal.

Interesting ideas

  • I1. Counts of recognizer output units (phone n-grams, words) are reweighted with tf.idf, so units that appear in nearly every utterance stop dominating the vector and each feature stays inspectable.
  • I2. The test data was deliberately recorded through a different channel than the training data, so a classifier cannot score well by memorizing recording quirks.
  • I3. Labeling used an adaptive stopping rule: start with a few judges and add more only until agreement crosses a threshold or a cap is reached.
  • I4. The classifier's confusions (dialect segments predicted as MSA) were compared with the confusions in human labels, which shows that some "errors" reflect genuine code-switching.
  • I5. A second network refines the first network's posteriors using a wider time window (31 frames), a cheap two-pass correction design.

For Mova

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

  • I1 (tree): The same transparent n-gram-over-labels vectors could detect which UD treebank dialect or genre a text resembles before choosing a conversion path.
  • I2 (learn): Evaluation sets could be drawn from a different source or genre than training text, serving as a control against shortcuts tied to corpus style.
  • I3 (learn): Silver data could call extra witness parsers or a higher LLM teacher tier only for sentences where the first witnesses disagree, saving cost on easy cases.
  • I4 (learn): Error analysis could check whether a parser mistake falls where human annotators also disagree, separating real bugs from inherently ambiguous input.
  • I5 (tree): gparse could add a light second pass that revises first-pass arc and label scores using wider sentence context, logging each change so it stays explainable.

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

🎓Johns Hopkins UniversityBaltimore · United States240 works · 96,487 citations🎓Massachusetts Institute of TechnologyCambridge · United States211 works · 69,000 citations🎓Hamad bin Khalifa UniversityDoha · Qatar141 works · 14,175 citations🎓École de Technologie SupérieureMontreal · Canada2 works · 60 citations

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

Cite

Cite as Ahmed Ali, Najim Dehak, Patrick Cardinal, Sameer Khurana, Sree Harsha Yella, James Glass, Peter Bell and Steve Renals. 2015. Automatic Dialect Detection in Arabic Broadcast Speech. arXiv:1509.06928.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.1509.06928 · arXiv 1509.06928

@misc{ali2015automatic,
  title = {{Automatic Dialect Detection in Arabic Broadcast Speech}},
  author = {Ahmed Ali, Najim Dehak, Patrick Cardinal, Sameer Khurana, Sree Harsha Yella, James Glass, Peter Bell and Steve Renals},
  year = {2015},
  doi = {10.48550/arxiv.1509.06928},
  eprint = {1509.06928},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1509.06928}
}

References

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

  1. G. Liu, Y. Lei, and J. H. Hansen, “Dialect identification: Impact of differences between read versus spontaneous speech,” EUSIPCO-2010: European Signal Processing Conference, Aalborg, Denmark, 2010. find
  2. D. A. Reynolds, W. M. Campbell, W. Shen, and E. Singer, “Automatic language recognition via spectral and token based approaches,” in Springer Handbook of Speech Processing, J. Benesty, M. M. Sondhi, and Y. Huang, Eds. Springer, 2008. ↗
  3. M. Zissman, “Comparison of four approaches to automatic language identification of telephone speech,” IEEE Transactions on Speech and Audio Processing, vol. 4, no. 1, pp. 31–44, 1996. ↗
  4. D. Martı́nez, L. Burget, L. Ferrer, and N. Scheffer, “ivector-based prosodic system for language identification,” in ICASSP, 2012, pp. 4861–4864. find
  5. O. Plchot, M. Diez, M. Soufifar, and L. Burget, “Pllr features in language recognition system for rats,” in Fifteenth Annual Conference of the International Speech Communication Association, 2014. find
  6. M. H. Bahari, N. Dehak, L. Burget, A. Ali, J. Glass et al., “Non-negative factor analysis for gmm weight adaptation,” IEEE Transactions on Audio Speech and Language Processing, 2014. find
  7. H. Soltau, L. Mangu, and F. Biadsy, “From modern standard arabic to levantine asr: Leveraging gale for dialects,” in ASRU, 2011, pp. 266–271. find
  8. M. Soufifar, S. Cumani, L. Burget, and J. Černocky, “Discriminative classifiers for phonotactic language recognition with ivectors,” in ICASSP, 2012, pp. 4853–4856. find
  9. A. Ali, Y. Zhang, P. Cardinal, N. Dahak, S. Vogel, and J. Glass, “A complete kaldi recipe for building arabic speech recognition systems,” in SLT, 2014. find
  10. G. Salton, a. Wong, and C. S. Yang, “A Vector Space Model for Automatic Indexing,” Magazine Communications of the ACM, vol. 18, no. 11, pp. 613–620, 1975. ↗
  11. W. Lowe, S. Mcdonald, W. Lowe, and S. Mcdonald, “Division of Informatics , University of Edinburgh Institute for Adaptive and Neural Computation The Direct Route : Mediated Priming in Semantic Space by The Direct Route : Mediated Priming in Semantic Space,” no. April, 2000. find
  12. S. Padó and M. Lapata, “Dependency-Based Construction of Semantic Space Models,” Computational Linguistics, vol. 33, no. December 2004, pp. 161–199, 2007. ↗
  13. N. Dehak, P. Kenny, R. Dehak, P. Dumouchel, and P. Ouellet, “Front-end factor analysis for speaker verification,” Audio, Speech, and Language Processing, IEEE Transactions on, pp. 788–798, 2011. ↗
  14. N. Dehak, P. A. Torres-Carrasquillo, D. A. Reynolds, and R. Dehak, “Language recognition via i-vectors and dimensionality reduction,” in INTERSPEECH, 2011, pp. 857–860. find
  15. P. Kenny, P. Ouellet, N. Dehak, V. Gupta, and P. Dumouchel, “A study of interspeaker variability in speaker verification,” Audio, Speech, and Language Processing, IEEE Transactions on, pp. 980–988, 2008. ↗
  16. A. Ali, Y. Zhang, and S. Vogel, “QCRI advanced transcription ssystem (QATS),” in SLT, 2014. find
  17. S. Wray and A. Ali, “Crowdsource a little to label a lot: Labeling a speech corpus of dialectal arabic,” in INTERSPEECH, 2015. find
  18. H. Drucker, D. Wu, and V. N. Vapnik, “Support vector machines for spam categorization,” Neural Networks, IEEE Transactions on, vol. 10, no. 5, pp. 1048–1054, 1999. ↗
  19. S. Meignier and T. Merlin, “Lium spkdiarization: an open source toolkit for diarization,” in CMU SPUD Workshop, 2010. find
  20. C.-h. Wu, H.-p. Shen, and C.-s. Hsu, “Code-Switching Event Detection by Using a Latent Language Space Model and the Delta-Bayesian,” vol. 23, no. 11, pp. 1892–1903, 2015. find
  21. P. Cardinal, N. Dehak, Y. Zhang, and J. Glass, “Speaker Adaptation Using the I-Vector Technique for Bottleneck Features,” in INTERSPEECH, 2015. find
  22. F. Richardson, D. Reynolds, and N. Dehak, “A Unified Deep Neural Network for Speaker and Language Recognition,” in INTER-SPEECH, 2015. find
  23. Y. Song, B. Jiang, Y. Bao, S. Wei, and L. R. Dai, “I-vector representation based on bottleneck features for language identification,” IEEE Electronics Letters, 2013. ↗
  24. P. Matejka, L. Zhang, T. Ng, H. S. Mallidi, O. Glembek, J. Ma, and B. Zhang, “Neural network bottleneck features for language identification,” in Proc. of Odyssey, 2014. This figure "DialectDetectionSpace.png" is available in "png" format from: http://arxiv.org/ps/1509.06928v2 This figure "bnf.png" is available in "png" format from: http://arxiv.org/ps/1509.06928v2 ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ali-2015-automatic-dialect-detection-arabic-broadcast/