--- title: "Arabic Little STT — Alkadri 2025" description: "Speech recognition systems are trained mostly on adult voices, and Arabic has no public corpus of children's speech." authors: ["Mouhand Alkadri", "Dania Desouki", "Khloud Al Jallad"] year: 2025 venue: "arXiv:2510.23319" source: "https://arxiv.org/abs/2510.23319" doi: "10.48550/arxiv.2510.23319" arxiv: "2510.23319" 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/alkadri-2025-arabic-little-stt-arabic-children/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/alkadri-2025-arabic-little-stt-arabic-children/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/) / Arabic Little STT: Arabic Children Speech Recognition Dataset # Arabic Little STT: Arabic Children Speech Recognition Dataset **Authors:** Mouhand Alkadri, Dania Desouki, [Khloud Al Jallad](https://md.research.elamur.ai/free/en/authors/khloud-al-jallad/) · **Year:** 2025 · **Venue:** arXiv:2510.23319 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://arxiv.org/abs/2510.23319) · [PDF](https://arxiv.org/pdf/2510.23319) · [DOI 10.48550/arxiv.2510.23319](https://doi.org/10.48550/arxiv.2510.23319) · [arXiv 2510.23319](https://arxiv.org/abs/2510.23319) · [OpenAlex](https://openalex.org/W7101950684) **Citations:** 0 citations (openalex-snapshot · 2026-10-14) · [1](#cited-by) in our library, of the 140884 papers whose bibliographies we read ## Summary Speech recognition systems are trained mostly on adult voices, and Arabic has no public corpus of children's speech. This paper fills that gap with Arabic Little STT, a small dataset of Syrian Levantine (Shami) child speech recorded in IT classrooms on ordinary smartphones. It contains 355 utterances from 288 children aged 6–13. The recordings are noisy, and the transcripts cover programming, robotics and AI. Native Shami speakers transcribed the audio by hand, following Whisper's conventions for non-English text, and an Arabic-specific normalization step was then applied. The authors tested eight Whisper variants without a language prompt and scored them with WER and CER. Every model struggled. The best, Large-v3, reached 66% WER, about four times the adult FLEURS figure the authors cite. The smallest models went far beyond 100% WER. The authors link this to similar findings for English children's speech. They conclude that models trained on adult data fail to generalize to children regardless of language or size, so dedicated child benchmarks, collected under strict ethical safeguards, are needed. ## Key points - Dataset: Arabic Little STT, released on Hugging Face (little-stt/little-stt-dataset). - 355 utterances from 288 children (157 boys, 131 girls), mostly aged 8–9. - Clips average about 10 s; transcripts average about 13 words (range 3–62). - Audio was recorded as AAC and converted to WAV. - Normalization applied to the transcripts: - removal of harakat and shadda; - removal of tatweel; - unification of alef variants (آ, إ, أ → ا); - removal of punctuation. - WER for the eight Whisper variants: | Model | WER | |---|---| | Tiny | 415% | | Base | 359% | | Small | 158% | | Medium | 119% | | Large | 82% | | Large-v2 | 82% | | Large-v3 | 66% (CER 34%) | | Large-v3 Turbo | 125% | - Language detection was very accurate: the Large variant identified Arabic correctly over 99% of the time. Transcription quality was still poor. - On the Open Universal Arabic ASR Leaderboard (August 2025), Whisper large-v3 ranks third with an average WER of 36.86. NVIDIA FastConformer models hold the top two places. ## Interesting ideas - The same normalization is applied to both the reference and the model output, so orthographic noise does not distort the error rate. — For Mova (learn): Mova's evaluation could canonicalize gold and predicted outputs (lemma spelling, punctuation tokens) identically before scoring, so differences in convention are not counted as errors. - Language-identification accuracy is reported separately from transcription accuracy. This shows that a correct early decision does not guarantee correct downstream output. — For Mova (tokens): accuracy could be reported per stage (tokenization, tagging, then lemmas), which shows where a pipeline actually fails rather than only that it fails. - The whole family of model sizes is evaluated, from Tiny to Large-v3 Turbo. This exposes non-monotonic surprises, such as the distilled Turbo doing much worse than its parent. — For Mova (tree): perceptron and gparse variants of different sizes could be benchmarked side by side, so a faster model is not assumed to be nearly as accurate. - The domain gap is measured by comparing scores on the new out-of-domain set with published in-domain baselines for the same models. — For Mova (tree): parsing accuracy on simple child-level or school texts could be compared with standard UD treebank scores to quantify the shift Mova faces in school-course material. - The test set is gathered under realistic deployment conditions (classroom noise, cheap microphones, one topic domain) rather than idealized ones. — For Mova (answer): test material for the readers could be drawn from real school tests and fables as students actually meet them, so the scores reflect true use. *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/) ## Cited by Cited by 1 paper of our library cites it (of the 140884 whose bibliographies we read); each has a page here. - [Abjad-Kids: An Arabic Speech Classification Dataset for Primary Education](https://md.research.elamur.ai/free/en/papers/snoubara-2026-abjad-kids-arabic-speech-classification/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @misc{alkadri2025arabic, title = {{Arabic Little STT: Arabic Children Speech Recognition Dataset}}, author = {Alkadri, Mouhand and Desouki, Dania and Jallad, Khloud Al}, year = {2025}, doi = {10.48550/arxiv.2510.23319}, eprint = {2510.23319}, archivePrefix = {arXiv}, url = {https://doi.org/10.48550/arxiv.2510.23319} } ``` - [← previous Arabic Large Language Models for Medical Text Generation](https://md.research.elamur.ai/free/en/papers/allam-2025-arabic-large-language-models-medical/) - [next → Arabic Metaphor Sentiment Classification Using Semantic Information](https://md.research.elamur.ai/free/en/papers/alsiyat-2025-arabic-metaphor-sentiment-classification-using/) - [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/) ## References 17 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. [2507.13977] Open Automatic Speech Recognition Models for Classical and Modern Standard Arabic. (n.d.). Retrieved August 16, 2025, from https://www.arxiv.org/abs/2507.13977 [↗](https://arxiv.org/abs/2507.13977) [our card](https://md.research.elamur.ai/free/en/papers/grigoryan-2025-open-automatic-speech-recognition-models/) 2. Ardila, R., Branson, M., Davis, K., Henretty, M., Kohler, M., Meyer, J., Morais, R., Saunders, L., Tyers, F. M., & Weber, G. (2020). Common Voice: A Massively-Multilingual Speech Corpus (No. arXiv:1912.06670). arXiv. https://doi.org/10.48550/arXiv.1912.06670 Conneau, A., Ma, M., Khanuja, S., Zhang, Y., Axelrod, V., Dalmia, S., Riesa, J., Rivera, C., & Bapna, A. [↗](https://doi.org/10.48550/arxiv.1912.06670) [our card](https://md.research.elamur.ai/free/en/papers/ardila-2019-common-voice-massively-multilingual-speech/) 3. (2022). FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech (No. arXiv:2205.12446). arXiv. https://doi.org/10.48550/arXiv.2205.12446 [↗](https://doi.org/10.48550/arxiv.2205.12446) [our card](https://md.research.elamur.ai/free/en/papers/conneau-2022-fleurs-few-shot-learning-evaluation/) 4. Hernandez, F., Nguyen, V., Ghannay, S., Tomashenko, N., & Estève, Y. (2018). TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation (Vol. 11096, pp. 198– 208). https://doi.org/10.1007/978-3-319-99579-3_21 [↗](https://doi.org/10.1007/978-3-319-99579-3_21) 5. Jain, R., Barcovschi, A., Yiwere, M., Corcoran, P., & Cucu, H. (2023). Adaptation of Whisper models to child speech recognition (No. arXiv:2307.13008). arXiv. https://doi.org/10.48550/arXiv.2307.13008 Librispeech: An ASR corpus based on public domain audio books | IEEE Conference Publication | IEEE Xplore. (n.d.). Retrieved May 4, 2025, from https://ieeexplore.ieee.org/document/7178964 [↗](https://doi.org/10.48550/arxiv.2307.13008) 6. (PDF) The PF_STAR children’s speech corpus. (n.d.). Retrieved May 3, 2025, from [https://www.researchgate.net/publication/221491189_The_PF_STAR_children's_speech_corpus](https://www.researchgate.net/publication/221491189_The_PF_STAR_children's_speech_corpus) [↗](https://www.researchgate.net/publication/221491189_The_PF_STAR_children's_speech_corpus) 7. Pradhan, S. S., Cole, R. A., & Ward, W. H. (2023). My Science Tutor (MyST)—A Large Corpus of Children’s Conversational Speech (No. arXiv:2309.13347). arXiv. https://doi.org/10.48550/arXiv.2309.13347 [↗](https://doi.org/10.48550/arxiv.2309.13347) 8. Puvvada, K. C., Żelasko, P., Huang, H., Hrinchuk, O., Koluguri, N. R., Dhawan, K., Majumdar, S., Rastorgueva, E., Chen, Z., Lavrukhin, V., Balam, J., & Ginsburg, B. (2024). Less is More: Accurate Speech Recognition & Translation without Web-Scale Data (No. arXiv:2406.19674). arXiv. https://doi.org/10.48550/arXiv.2406.19674 [↗](https://doi.org/10.48550/arxiv.2406.19674) 9. Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., & Sutskever, I. (n.d.). Robust Speech Recognition via Large-Scale Weak Supervision. arXiv.Org. Retrieved May 3, 2025, from https://arxiv.org/abs/2212.04356v1 [↗](https://arxiv.org/abs/2212.04356) [our card](https://md.research.elamur.ai/free/en/papers/radford-2022-robust-speech-recognition-large-scale/) 10. Rousseau, A., Deléglise, P., & Estève, Y. (2012). TED-LIUM: An Automatic Speech Recognition dedicated corpus. In N. Calzolari, K. Choukri, T. Declerck, M. U. Doğan, B. Maegaard, J. Mariani, A. Moreno, J. Odijk, & S. Piperidis (Eds.), Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC`12) (pp. 125–129). European Language Resources Association (ELRA). https://aclanthology.org/L12-1405/ [↗](https://aclanthology.org/L12-1405/) 11. Rousseau, A., Deléglise, P., & Estève, Y. (2014). Enhancing the TED-LIUM Corpus with Selected Data for Language Modeling and More TED Talks. In N. Calzolari, K. Choukri, T. Declerck, H. Loftsson, B. Maegaard, J. Mariani, A. Moreno, J. Odijk, & S. Piperidis (Eds.), Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC`14) (pp. 3935– 3939). European Language Resources Association (ELRA). https://aclanthology.org/L14-1079/ [↗](https://aclanthology.org/L14-1079/) 12. Schneider, S., Baevski, A., Collobert, R., & Auli, M. (2019). wav2vec: Unsupervised Pre-training for Speech Recognition (No. arXiv:1904.05862). arXiv. https://doi.org/10.48550/arXiv.1904.05862 [↗](https://doi.org/10.48550/arxiv.1904.05862) 13. Segal-Feldman, Y., Shamsian, A., Navon, A., Hetz, G., & Keshet, J. (2024). Whisper in Medusa’s Ear: Multi-head Efficient Decoding for Transformer-based ASR (No. arXiv:2409.15869). arXiv. https://doi.org/10.48550/arXiv.2409.15869 [↗](https://doi.org/10.48550/arxiv.2409.15869) 14. Shi, Z., Srivastava, H., Shi, X., Narayanan, S., & Matarić, M. J. (2024). Personalized Speech Recognition for Children with Test-Time Adaptation (No. arXiv:2409.13095). arXiv. https://doi.org/10.48550/arXiv.2409.13095 [↗](https://doi.org/10.48550/arxiv.2409.13095) 15. Srivastav, V., Majumdar, S., Koluguri, N., Moumen, A., Gandhi, S., & others. (2023). Open Automatic Speech Recognition Leaderboard. Hugging Face. https://huggingface.co/spaces/hf-audio/open_asr_leaderboard Talafha, B., Kadaoui, K., Magdy, S. M., Habiboullah, M., Chafei, C. M., El-Shangiti, A. O., Zayed, H., tourad, M. cheikh, Alhamouri, R., Assi, R., Alraeesi, A., Mohamed, H., Alwajih, F., Mohamed, A., Mekki, A. E., Nagoudi, E. M. B., Saadia, B. D. M., Alsayadi, H. A., Al-Dhabyani, W., … Abdul-Mageed, M. (2024). Casablanca: Data and Models for Multidialectal Arabic Speech Recognition (No. arXiv:2410.04527). arXiv. https://doi.org/10.48550/arXiv.2410.04527 [↗](https://doi.org/10.48550/arxiv.2410.04527) 16. The CMU Kids Corpus—Linguistic Data Consortium. (n.d.). Retrieved May 3, 2025, from [https://catalog.ldc.upenn.edu/LDC97S63](https://catalog.ldc.upenn.edu/LDC97S63) [↗](https://catalog.ldc.upenn.edu/LDC97S63) 17. Wang, Y., Alhmoud, A., & Alqurishi, M. (2024). Open Universal Arabic ASR Leaderboard. arXiv Preprint arXiv:2412.13788. [↗](https://arxiv.org/abs/2412.13788) [our card](https://md.research.elamur.ai/free/en/papers/wang-2024-open-universal-arabic-asr-leaderboard/)