Elamur Research⌕
☰
home / papers / 2025 / Hybrid Deep Learning and Signal Processing for Arabic Dialect Recognition in Low-Resource Settings

Hybrid Deep Learning and Signal Processing for Arabic Dialect Recognition in Low-Resource Settings

Summary

Identifying which Arabic dialect a speaker uses is difficult: dialects differ in sound, vocabulary and grammar, and most have little labeled audio. The authors test whether classical signal-processing features paired with neural classifiers can work when only a few hours of data exist. They built a three-way dataset (Egyptian, Levantine, Gulf) from the Arabic part of Common Voice 12.0, using the speaker's stated country as the dialect label. After cleaning, resampling and filtering, about six balanced hours remained, and training was supported by augmentation. Two main pipelines were compared: MFCC features fed to a CNN, and Discrete Wavelet Transform coefficients fed to an RNN. The crossed pairings were also run to separate the effect of the features from the effect of the classifier. MFCC + CNN reached 91.2% accuracy, far above the 66.5% of Wavelet + RNN. The crossed runs suggest the MFCC representation accounts for most of the gain. The result is a cheap, reproducible baseline for dialect identification before moving to self-supervised or Transformer models.

Key points

  • Data: the Arabic subset of Mozilla Common Voice 12.0, loaded through Hugging Face Datasets. Clips were mapped to three dialect groups by country metadata, giving about 6 hours split evenly. Audio was converted to mono WAV at 16 kHz and normalized.
  • Features: MFCCs were computed with Librosa (STFT with Hamming window, mel filterbank, log, DCT). The DWT used PyWavelets with Daubechies-4 decomposed to level 3.
  • Models: the CNN has three convolutional layers with max-pooling, a dense layer and a softmax output. The RNN, built in PyTorch, uses SimpleRNN or LSTM layers depending on tuning.
  • Results: MFCC + CNN scored 91.2% accuracy, 92.8% precision, 91.2% recall and 91.0% F1. Wavelet + RNN scored 66.5% accuracy. For comparison, earlier CNN work by Shon et al. on MGB-3 reported up to 78%.
  • Augmentation and limitations: augmentation used pitch shift, time stretch, noise injection and speed perturbation. Stated limitations are the small dataset, possible accent overlap under country-based labels, and an under-tuned wavelet model.

Interesting ideas

  • I1. Instead of only comparing two finished pipelines, the authors cross every feature type with every classifier and then assign each factor a role (major driver, secondary, weak); this factorial ablation shows what actually causes a gain.
  • I2. Labels are derived cheaply from existing metadata (speaker country) rather than hand annotation, with the authors openly noting that this proxy blurs class boundaries.
  • I3. The paper attributes the wavelet model's failure partly to flattening its features, which discards the structure a sequential model needs; how the input is laid out matters as much as what it contains.
  • I4. A compact, well-understood hand-designed representation combined with a modest learner beat a richer but less structured one in a low-data regime.

For Mova

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

  • I1 (learn): Evaluation could cross input representations (e.g., tokenizer variants) with parsers (perceptron vs gparse) to attribute LAS changes to the right component instead of the whole pipeline.
  • I2 (learn): Silver labels taken from source metadata, such as genre or register of a text, can bootstrap training as long as a negative-control check measures how much label noise the proxy introduces.
  • I3 (tree): Parser features should keep the tree and sequence structure (head, dependents, order) rather than squashing them into flat vectors, so that the model stays explainable and not bag-of-words.
  • I4 (tokens): For rare morphology or lemmas with few examples, explicit engineered cues such as suffix and shape features may outperform heavier learned representations until more data exists.

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: Linguistics & languages

Topics Low-resource languages Semitic languages

Cite

Cite as Ghazal Al-Shwayyat and Omer Nezih Gerek. 2025. Hybrid Deep Learning and Signal Processing for Arabic Dialect Recognition in Low-Resource Settings. arXiv:2506.21386.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2506.21386 · arXiv 2506.21386

@misc{alshwayyat2025hybrid,
  title = {{Hybrid Deep Learning and Signal Processing for Arabic Dialect Recognition in Low-Resource Settings}},
  author = {Ghazal Al-Shwayyat and Omer Nezih Gerek},
  year = {2025},
  doi = {10.48550/arxiv.2506.21386},
  eprint = {2506.21386},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2506.21386}
}

References

22 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Open speech and language resources (openslr), 2015. https://openslr.org. ↗
  2. Martı́n Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mane, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viegas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. Tensorflow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org. find
  3. Osama Abdel-Hamid, Abdel-rahman Mohamed, Hui Jiang, Li Deng, Gerald Penn, and Dong Yu. Convolutional neural networks for speech recognition. In IEEE/ACM Transactions on Audio, Speech, and Language Processing, volume 22, pages 1533–1545. IEEE, 2014. find
  4. Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Reuben Henretty, Michael Morais, Lindsay Saunders, Francis M. Tyers, and Gregor Weber. Common voice: A massively-multilingual speech corpus. Proceedings of The 12th Language Resources and Evaluation Conference (LREC), pages 4218–4222, 2020. our card
  5. Steven Davis and Paul Mermelstein. Comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences. IEEE Transactions on Acoustics, Speech, and Signal Processing, 28(4):357–366, 1980. find
  6. Mona Diab, Nizar Habash, and Imed Zitouni. Nlp for arabic and related languages. AI Magazine, 58(3):9–13, 2018. find
  7. A. Djanibekov, H. O. Toyin, R. Alshalan, A. Alitr, and H. Aldarmaki. Dialectal coverage and generalization in arabic speech recognition. 2024. Preprint available at https://doi.org/10.48550/arxiv.2411.05872. ↗
  8. Youssef Fares, Zeinab El-Zanaty, Khaled Abdel-Salam, Mohamed Ezzeldin, Ahmed Mohamed, Karim El-Awaad, and Mustafa Torki. Arabic dialect identification with deep learning and hybrid frequency based features. In Proceedings of the Meeting of the Association for Computational Linguistics, pages 224–228, 2019. find
  9. Geoffrey Hinton, Li Deng, Dong Yu, George E. Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N. Sainath, and Brian Kingsbury. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal Processing Magazine, 29(6):82–97, 2012. find
  10. Sepp Hochreiter and Jürgen Schmidhuber. 9(8):1735–1780, 1997. find
  11. Tom Ko, Vijayaditya Peddinti, Daniel Povey, Michael L. Seltzer, and Sanjeev Khudanpur. Audio augmentation for speech recognition. In Proceedings of Interspeech, pages 3586–3589, 2015. find
  12. Gregory Lee et al. Pywavelets: A python package for wavelet analysis. Journal of Open Source Software, 4(36):1237, 2019. find
  13. Stéphane Mallat. A Wavelet Tour of Signal Processing. Academic Press, 1999. find
  14. Brian McFee, Colin Raffel, Dawen Liang, Daniel PW Ellis, Matt McVicar, Eric Battenberg, and Oriol Nieto. librosa: Audio and music signal analysis in python. In Proceedings of the 14th Python in Science Conference (SciPy 2015), pages 18–25, 2015. find
  15. T. M. Nazmy, M. E. Gadallah, and A. A. Abdelhamid. A novel method for arabic consonant/vowel segmentation using wavelet transform. Egyptian Computer Science Journal, 27, 2005. find
  16. Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V. Le. Specaugment: A simple data augmentation method for automatic speech recognition. In Proceedings of Interspeech, pages 2613–2617, 2019. our card
  17. Adam Paszke et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, volume 32, pages 8026–8037, 2019. our card
  18. Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Matthieu Perrot, and Édouard Duchesnay. Scikit-learn: Machine learning in python. Journal of Machine Learning Research, 12:2825–2830, 2011. find
  19. David M. W. Powers. Evaluation: From precision, recall and f-measure to roc, informedness, markedness and correlation. Journal of Machine Learning Technologies, 2(1):37–63, 2011. find
  20. A. Rahman, Md. M. Kabir, M. F. Mridha, M. Alatiyyah, H. F. Alhasson, and S. S. Alharbi. Arabic speech recognition: Advancement and challenges. IEEE Access, 2024. find
  21. Sangwon Shon, Amr Ali, and James Glass. Convolutional neural networks and language embeddings for end-to-end dialect recognition. In Proceedings of the Odyssey: The Speaker and Language Recognition Workshop, pages 98–104, 2018. find
  22. Ghazal Shwayat. Arabic dialect recognition hybrid models (code repository). https://github.com/Ghazi444/arabic-dialect-recognition-hybrid, 2025. Accessed June 2025. Long short-term memory. Neural Computation, ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/al-shwayyat-2025-hybrid-deep-learning-signal-processing/