---
title: "Abjad-Kids: An Arabic Speech Classification — Snoubara 2026"
description: "Child speech is still poorly covered by public speech data, and Arabic is especially underserved."
authors: ["Abdul Aziz Snoubara", "Baraa Al_Maradni", "Haya Al_Naal", "Malek Al_Madrmani", "Roaa Jdini", "Seedra Zarzour", "Khloud Al Jallad"]
year: 2026
venue: "arXiv:2603.20255"
source: "https://arxiv.org/abs/2603.20255"
doi: "10.48550/arxiv.2603.20255"
arxiv: "2603.20255"
licence: "CC BY 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/snoubara-2026-abjad-kids-arabic-speech-classification/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/snoubara-2026-abjad-kids-arabic-speech-classification/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/3/) / Abjad-Kids: An Arabic Speech Classification Dataset for Primary Education

# Abjad-Kids: An Arabic Speech Classification Dataset for Primary Education

**Authors:** Abdul Aziz Snoubara, Baraa Al\_Maradni, Haya Al\_Naal, Malek Al\_Madrmani, Roaa Jdini, Seedra Zarzour, [Khloud Al Jallad](https://md.research.elamur.ai/free/en/authors/khloud-al-jallad/) · **Year:** 2026 · **Venue:** arXiv:2603.20255 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2603.20255) · [PDF](https://arxiv.org/pdf/2603.20255) · [DOI 10.48550/arxiv.2603.20255](https://doi.org/10.48550/arxiv.2603.20255) · [arXiv 2603.20255](https://arxiv.org/abs/2603.20255) · [OpenAlex](https://openalex.org/W7140166013)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
Child speech is still poorly covered by public speech data, and Arabic is especially underserved. Models trained on adult voices do not transfer well to children, and many Arabic letters sound alike, which makes the problem harder. The authors build Abjad-Kids, a classification dataset for early schooling. It contains 46,397 recordings from children aged 3–12, spanning 141 classes of letters, numbers and colors. All recordings follow fixed technical specifications for length, sampling rate and file format.

For alphabet recognition they use a two-step CNN-LSTM pipeline. A first model assigns a recording to a group of letters, and a dedicated classifier then picks the letter within that group. Two ways of forming groups were compared: groups fixed in advance from articulation points, and groups found by clustering. The linguistically motivated groups worked better. The proposed models beat a baseline architecture and classical machine-learning methods, and augmentation improved results. However, train and test accuracy differ markedly, which points to too little data per class. The resource is meant to support Arabic educational speech tools for children.

## Key points
- Abjad-Kids contains 46,397 audio clips from speakers aged 3–12, covering 141 classes (alphabet letters, numbers, colors), recorded under uniform specifications.
- Alphabet recognition is split into a group-level classifier followed by per-group CNN-LSTM classifiers. Static articulation-based grouping (e.g., a Halq group) outperformed dynamic clustering-based grouping.
- Many models exceed 95% train accuracy while test accuracy stays between 80% and 92%. In the Halq group, train accuracy reached 99%, a sign of memorization.
- On separate tasks, the best color model reached 83% test accuracy. The best number model reached 86% test accuracy, against 80% for the baseline architecture. The train–test gap ranges from 5% to 14%.
- Pitch shifting, low-pass filtering and gain adjustment raised test accuracy on all tasks, most of all for small letter groups. Dropout of 0.1–0.5 limited overfitting but did not remove it.

## Interesting ideas
- Breaking a large confusable label set into expert-defined coarse groups, then training a specialist per group, reduced confusion between similar classes better than one flat classifier. — For Mova (tokens): The tagger could first predict a coarse class, such as verb-like versus nominal, and then run specialised classifiers for fine morphological features within each class, so confusable tags compete only with their true neighbours.
- The paper compares groupings built from linguistic knowledge against groupings learned by clustering on the same task, and the knowledge-based ones won. — For Mova (frames): When grouping verbs for role assignment, test VerbNet classes against clusters induced from parser output on the same benchmark, to check whether hand-built classes really earn their place.
- Train and test accuracy are reported separately for every subgroup, which exposes exactly where a model memorises rather than generalises. — For Mova (learn): Evaluation could report the train–dev gap per dependency relation or frame type, so overfitting on rare categories is flagged instead of being hidden in a global score.
- Augmentation was most useful for the groups with the fewest examples per class, so targeting scarce classes pays off more than uniform augmentation. — For Mova (learn): Silver-data generation could be steered toward rare relations and constructions, with extra agreed examples produced where the training set is thinnest.
- Recordings were kept technically uniform, and a cited study showed that classifiers can latch onto recording artifacts instead of real features. — For Mova (learn): Negative controls should check that the parser is not keying on treebank-specific annotation quirks, for example by testing on text converted from another UD dialect.

*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](https://md.research.elamur.ai/free/en/shelves/linguistics/)

Topics: [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/)

## Cite

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

```bibtex
@misc{snoubara2026abjad,
  title = {{Abjad-Kids: An Arabic Speech Classification Dataset for Primary Education}},
  author = {Snoubara, Abdul Aziz and Al\_Maradni, Baraa and Al\_Naal, Haya and Al\_Madrmani, Malek and Jdini, Roaa and Zarzour, Seedra and Jallad, Khloud Al},
  year = {2026},
  doi = {10.48550/arxiv.2603.20255},
  eprint = {2603.20255},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2603.20255}
}
```

- [← previous A Transparent Model of Syntactic and Semantic Cue-based Retrieval](https://md.research.elamur.ai/free/en/papers/yue-2026-transparent-model-syntactic-semantic-cue/)
- [next → Abraham Firkowitsh, Karaites in Hit, and the Provenance of Karaite Transcription of the Biblical Hebrew Texts into Arabic Script](https://md.research.elamur.ai/free/en/papers/harviainen-2026-abraham-firkowitsh-karaites-hit-provenance/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/3/)

## References

19 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. Yeung and A. Alwan, On the Difficulties of Automatic Speech Recognition for Kindergarten-Aged Children. 2018. doi: 10.21437/Interspeech.2018-2297. [↗](https://doi.org/10.21437/interspeech.2018-2297)
2. A. A. Attia, J. Liu, W. Ai, D. Demszky, and C. Espy-Wilson, “Kid-Whisper: Towards Bridging the Performance Gap in Automatic Speech Recognition for Children VS. Adults,” Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, vol. 7, no. 1, pp. 74–80, Oct. 2024, doi: 10.1609/aies.v7i1.31618. [↗](https://doi.org/10.1609/aies.v7i1.31618)
3. F. Yu et al., “The SLT 2021 children speech recognition challenge: Open datasets, rules and baselines,” 2020. [Online]. Available: https://arxiv.org/abs/2011.06724 [↗](https://arxiv.org/abs/2011.06724)
4. L. Fahkrunisa, I. Sulistiawati, and R. M, “CHILDREN’S INTERACTION WITH AI TECHNOLOGY AND ITS IMPACT ON LANGUAGE ACQUISITION IN THE DIGITALIZATION ERA,” IJEE (Indonesian Journal of English Education), vol. 12, Mar. 2025, doi: 10.15408/ijee.v12i2.48667. [↗](https://doi.org/10.15408/ijee.v12i2.48667)
5. M. Alkadri, D. Desouki, and K. Al Jallad, “Arabic Little STT: Arabic Children Speech Recognition Dataset,” 2025. [Online]. Available: https://arxiv.org/abs/2510.23319 [↗](https://arxiv.org/abs/2510.23319) [our card](https://md.research.elamur.ai/free/en/papers/alkadri-2025-arabic-little-stt-arabic-children/)
6. S. Dutta et al., “Challenges remain in Building ASR for Spontaneous Preschool Children Speech in Naturalistic Educational Environments,” in Interspeech, 2022. [Online]. Available: [https://api.semanticscholar.org/CorpusID:252343724](https://api.semanticscholar.org/CorpusID:252343724) [↗](https://api.semanticscholar.org/CorpusID:252343724)
7. P. Shivakumar and P. Georgiou, Transfer Learning from Adult to Children for Speech Recognition: Evaluation, Analysis and Recommendations. 2018. doi: 10.48550/arXiv.1805.03322. [↗](https://doi.org/10.48550/arxiv.1805.03322)
8. R. Gary Leonard and George Doddington, “TIDIGITS LDC93S10,” Linguistic Data Consortium. Accessed: Mar. 10, 2026. [Online]. Available: [https://catalog.ldc.upenn.edu/LDC93S10](https://catalog.ldc.upenn.edu/LDC93S10) [↗](https://catalog.ldc.upenn.edu/LDC93S10)
9. J. Mostow, “KIDS: A database of children’s speech,” Journal of the Acoustical Society of America, vol. 100, no. 4, Oct. 1996. [find](https://scholar.google.com/scholar?q=J%2E+Mostow%2C+%E2%80%9CKIDS%3A+A+database+of+children%E2%80%99s+speech%2C%E2%80%9D+Journal+of+the+Acoustical+Society+of+America%2C+vol%2E+100%2C+no%2E+4%2C+Oct%2E+1996%2E)
10. M. S. Atkins, D. J. Bailey, and S. Boyce, “Speech exemplar and evaluation database (SEED) for clinical training in articulatory phonetics and speech science,” Clin. Linguist. Phon., vol. 34, pp. 878–886, 2020, [Online]. Available: [https://api.semanticscholar.org/CorpusID:214616294](https://api.semanticscholar.org/CorpusID:214616294) [↗](https://api.semanticscholar.org/CorpusID:214616294)
11. K. Shobaki, J.-P. Hosom, and R. Cole, The OGI kids’ speech corpus and recognizers. 2000. doi: 10.21437/ICSLP.2000-800. [↗](https://doi.org/10.21437/icslp.2000-800)
12. S. Pradhan, R. A. Cole, and W. H. Ward, “My Science Tutor (MyST)–a Large Corpus of Children’s Conversational Speech,” in International Conference on Language Resources and Evaluation, 2023. [Online]. Available: [https://api.semanticscholar.org/CorpusID:262458961](https://api.semanticscholar.org/CorpusID:262458961) [↗](https://api.semanticscholar.org/CorpusID:262458961)
13. J. Zhou et al., “ChildMandarin: A Comprehensive Mandarin Speech Dataset for Young Children Aged 3-5,” in Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar, Eds., Vienna, Austria: Association for Computational Linguistics, Jul. 2025, pp. 12524–12537. doi: 10.18653/v1/2025.acl-long.614. [↗](https://doi.org/10.18653/v1/2025.acl-long.614)
14. T. Rolland and A. Abad, “Improved Children’s Automatic Speech Recognition Combining Adapters and Synthetic Data Augmentation,” ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 12757–12761, 2024, [Online]. Available: [https://api.semanticscholar.org/CorpusID:268525250](https://api.semanticscholar.org/CorpusID:268525250) [↗](https://api.semanticscholar.org/CorpusID:268525250)
15. R. Fan, N. Balaji Shankar, and A. Alwan, Benchmarking Children’s ASR with Supervised and Self-supervised Speech Foundation Models. 2024. doi: 10.21437/Interspeech.2024-1353. [↗](https://doi.org/10.21437/interspeech.2024-1353)
16. R. Fan, Y. Zhu, J. Wang, and A. Alwan, “Towards Better Domain Adaptation for Self-Supervised Models: A Case Study of Child ASR,” IEEE J. Sel. Top. Signal Process., vol. 16, pp. 1242–1252, 2022, [Online]. Available: [https://api.semanticscholar.org/CorpusID:251755792](https://api.semanticscholar.org/CorpusID:251755792) [↗](https://api.semanticscholar.org/CorpusID:251755792)
17. P. Bijith, A. Liu, L. Collomb, A. Gardiner, and S. Farrington, “Robust Child Speech Classification Leveraging Augmentation and Speaker Embeddings,” in 2025 IEEE 12th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), 2025, pp. 1212–1216. doi: 10.1109/ITAIC64559.2025.11163088. [↗](https://doi.org/10.1109/itaic64559.2025.11163088)
18. P. Ye and H. Duan, “A New Approach for Speech Keyword Spotting in Noisy Environment,” Mar. 2022, pp. 65–76. doi: 10.5121/csit.2022.120206. [↗](https://doi.org/10.5121/csit.2022.120206)
19. X. Liu, H. Lu, J. Yuan, and X. Li, “CAT: Causal Audio Transformer for Audio Classification,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5. doi: 10.1109/ICASSP49357.2023.10096787. [↗](https://doi.org/10.1109/icassp49357.2023.10096787)