Elamur Research⌕
☰
home / papers / 2025 / Tachelhiyt-Darija: a parallel speech corpus for two underrepresented languages

Tachelhiyt-Darija: a parallel speech corpus for two underrepresented languages

Paper · 2025 · Noureddine Atouf, Elsayed Issa and Said Ouzbayr

Summary

Speech technology still serves Amazigh and Moroccan Arabic (Darija) poorly, mainly because there is so little recorded, transcribed data for them. The authors build what they describe as the first parallel speech corpus pairing Tachelhiyt, an Amazigh variety, with Darija. They wrote or collected Darija sentences, partly from a Peace Corps textbook. Two native speakers then translated them into Tachelhiyt, and both sides were written in Arabic script rather than Tifinagh or phonetic notation, to make them easier to read. Native speakers recorded the sentences at home on their own devices. The recordings were split automatically, checked by hand and released on Hugging Face with speaker and gender metadata. As a baseline, the authors ran Whisper without any adaptation, and the error rates were very high, above 100% WER for Tachelhiyt. After fine-tuning Whisper small separately for each language, errors fell to single-digit WER. The work matters because it supplies a sentence-level, aligned resource for two neglected North African languages. It also shows that off-the-shelf multilingual ASR does little for them until it is adapted to local data.

Key points

  • The corpus has 2,772 segments recorded by 24 Tachelhiyt and 11 Darija speakers (35 in total, 60% female, mostly under 19). It contains 71.68 min of Tachelhiyt audio (average segment 3.11 s) and 61.84 min of Darija (average 2.68 s). It is published as NoureddineMOR/tachelhiyt-darija.
  • Audio was converted to mono 16 kHz 16-bit PCM WAV with FFmpeg. For each language there is a CSV of file name and transcript, plus an eight-column parallel metadata file with audio, transcript, speaker and gender for each side.
  • Zero-shot Whisper was tested on 300 samples (WER/CER):
    • Tachelhiyt: small 127.09/80.85, large-v3 145.54/85.86, so the larger model did worse.
    • Darija: small 89.72/60.47, large-v3 76.08/34.44.
  • Fine-tuning Whisper small used batch 16, learning rate 1e-5 and 1000 steps (about 15 epochs) on one A100. It reached 7.45/3.25 on Tachelhiyt and 4.26/1.38 on Darija.
  • The limitations the authors list are recording-quality variation from home devices and the small, parallel nature of the data. They plan an expanded corpus covering all three Tamazight varieties.

Interesting ideas

  • I1. Pivot-language corpus building: sentences are written first in the better-resourced variety, then translated by native speakers into the scarcer one, so alignment exists from the start without needing scarce source texts.
  • I2. Built-in segment boundaries: recorders left a fixed five-second pause between items, so splitting could be automatic, and a person only had to check the alignment afterwards.
  • I3. Scaling as a test condition, not an assumption: comparing two model sizes zero-shot showed the bigger Whisper getting worse on the unseen language while improving on the other one.
  • I4. One shared script for both sides: writing both languages in Arabic script makes the pair easier to read and compare, instead of mixing Tifinagh and phonetic notations.
  • I5. Parallel metadata with speaker and gender fields: this lets users filter and break down results by subgroup and by language side for each aligned item.

For Mova

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

  • I1 (back): build paired English source/target sets this way for testing translation through the meta-world, with meaning fixed before the second wording is produced.
  • I2 (learn): when collecting silver data, add explicit delimiters or IDs at creation time, so alignment is checked rather than guessed.
  • I3 (tree): compare gparse against the perceptron parser on out-of-domain or dialectal text, and do not assume the larger model wins everywhere.
  • I4 (tokens): map spelling and dialect variants into one canonical orthography before tagging and lemmatization, so the same word is not split across forms.
  • I5 (answer): attach source, variant and paraphrase-pair IDs to test items, so errors on school tests or fable QA can be sliced by subgroup.

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, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 works

🧭 Topics 1Research subjects and languages the work is about.Speech & spoken language · 519 works

Cite

Cite as Noureddine Atouf, Elsayed Issa and Said Ouzbayr. 2025. Tachelhiyt-Darija: a parallel speech corpus for two underrepresented languages. Proceedings of the 8th International Conference on Natural Language and Speech Processing (ICNLSP-2025).

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

@inproceedings{atouf2025tachelhiyt,
  title = {{Tachelhiyt-Darija: a parallel speech corpus for two underrepresented languages}},
  author = {Noureddine Atouf, Elsayed Issa and Said Ouzbayr},
  booktitle = {Proceedings of the 8th International Conference on Natural Language and Speech Processing (ICNLSP-2025)},
  year = {2025},
  url = {https://aclanthology.org/2025.icnlsp-1.37/}
}

References

22 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Fadwa Abakarim and Abdenbi Abenaou. 2023. Enhancing amazigh speech recognition system with mfdwc-svm. In International Conference on Computational Science and Its Applications, pages 471–488. Springer. find
  2. Abdelilah El Aissati, Susy Karsmakers, and Jeanne Kurvers. 2011. “we are all beginners”: Amazigh in language policy and educational practice in morocco. Compare: A Journal of Comparative and International Education, 41(2):211–227. find
  3. Ahmed Ali, Suwon Shon, Younes Samih, Hamdy Mubarak, Ahmed Abdelali, James Glass, Steve Renals, and Khalid Choukri. 2019. The mgb-5 challenge: Recognition and dialect identification of dialectal arabic speech. In 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 1026–1033. IEEE. find
  4. Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henretty, Reuben Morais, Lindsay Saunders, Francis Tyers, and Gregor Weber. 2020. Common voice: A massively-multilingual speech corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4218–4222, Marseille, France. European Language Resources Association.
  5. Hossam Boulal, Mohamed Hamidi, Mustapha Abarkan, and Jamal Barkani. 2023. Amazigh spoken digit recognition using a deep learning approach based on mfcc. International journal of electrical and computer engineering systems, 14(7):791–798. find
  6. Safâa El Ouahabi, Mohamed Atounti, and Mohamed Bellouki. 2017. A database for amazigh speech recognition research: Amzsrd. In 2017 3rd International Conference of Cloud Computing Technologies and Applications (CloudTech), pages 1–5. IEEE. find
  7. Moha Ennaji. 2014. Multiculturalism and Democracy in North Africa: Aftermath of the Arab Spring. Routledge. find
  8. Mohamed Hamidi, Hassan Satori, Ouissam Zealouk, and Khalid Satori. 2020. Amazigh digits through interactive speech recognition system in noisy environment. International Journal of Speech Technology, 23(1):101–109. find
  9. High Commission for Planning. 2024. Morocco population and housing census. find
  10. Maria Labied, Abdessamad Belangour, and Mouad Banane. 2023. Darija-c: towards a moroccan darija speech recognition and speech-to-text translation corpus. In 2023 1st International Conference on Advanced Innovations in Smart Cities (ICAISC), pages 1–4. IEEE. find
  11. Nourredine Oukas, Tiziri Chabi, and Tilelli Sari. 2024. A novel dataset for arabic speech recognition recorded by tamazight speakers. Authorea Preprints. find
  12. Aissam Outchakoucht and Hamza Es-Samaali. 2021. Moroccan dialect-darija-open dataset. arXiv preprint arXiv:2103.09687. ↗
  13. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492–28518. PMLR.
  14. Fatima Sadiqi. 2014. Berber and language politics in the moroccan educational system. In Moha Ennaji, editor, Multiculturalism and Democracy in North Africa: Aftermath of the Arab Spring. Routledge. find
  15. Younes Samih and Wolfgang Maier. 2016. An arabic-moroccan darija code-switched corpus. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16), pages 4170–4175.
  16. Hassan Satori and Fatima ElHaoussi. 2014. Investigation amazigh speech recognition using cmu tools. International Journal of Speech Technology, 17:235– 243. find
  17. Bashar Talafha, Karima Kadaoui, Samar Mohamed Magdy, Mariem Habiboullah, Chafei Mohamed Chafei, Ahmed Oumar El-Shangiti, Hiba Zayed, Rahaf Alhamouri, Rwaa Assi, Aisha Alraeesi, et al. 2024. Casablanca: Data and models for multidialectal arabic speech recognition. arXiv preprint arXiv:2410.04527. ↗
  18. Meryam Telmem and Youssef Ghanou. 2018. Estimation of the optimal hmm parameters for amazigh speech recognition system using cmu-sphinx. Procedia Computer Science, 127:92–101. find
  19. Suramya Tomar. 2006. Converting video formats with ffmpeg. Linux Journal, 2006(146):10. find
  20. Abderrahim Youssi. 1992. Grammaire et lexique de l’arabe marocain moderne. Wallada, Casablanca. find
  21. Hajar Zaidani, Abderrahim Maizate, Mohammed Ouzzif, and Rim Koulali. 2024a. Building a corpus for the underexplored moroccan dialect (cfmd) through audio segmentations. Revue d’Intelligence Artificielle, 38(3):857. find
  22. Hajar Zaidani, Abderrahim Maizate, Mohammed Ouzzif, and Rim Koulali. 2024b. Cfmd: Corpus for moroccan dialect as under researched dialect. In Future of Information and Communication Conference, pages 61–69. Springer. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/atouf-2025-tachelhiyt-darija-parallel-speech-corpus/