Elamur Research⌕
☰
home / papers / 2025 / TuniFra: A Tunisian Arabic Speech Corpus with Orthographic Transcriptions and French Translations

TuniFra: A Tunisian Arabic Speech Corpus with Orthographic Transcriptions and French Translations

Summary

Tunisian Arabic has very little speech data, and existing Tunisian corpora offer no French translations. The authors address this with TuniFra, which they describe as the first publicly available three-way resource for this dialect: audio, orthographic transcription, and French translation. They took 19 Tunisian YouTube podcasts (interviews and multi-speaker debates on topics such as ecology, education and economy), about 15 hours in total. Native Tunisian linguists segmented and transcribed the audio in Transcriber and then translated it into French. They followed explicit spelling conventions for frequent switching into French and English. As baselines, they fine-tuned Whisper models and a Wav2Vec-BERT CTC model for ASR. For speech translation, they compared cascades (ASR followed by NLLB) with an end-to-end model that joins a Whisper encoder to NLLB through a small CNN bridge. Fine-tuning cut WER sharply compared with zero-shot use. Wav2Vec-BERT gave the best ASR result, and cascades beat end-to-end translation. The corpus fills a gap for dialectal ASR and STT research.

Key points

  • TuniFra corpus: 15 h of speech, 9,189 segments averaging about 5.9 s, and 41 speakers (29 male, 8 female, 4 unknown). Of the segments, 5,353 are pure Tunisian, 132 are fully foreign and 3,704 are mixed. Overlapping speech is reported in 2,710 segments.
  • Annotation was done by two linguists for transcription and four for translation. Inter- and intra-annotator agreement was not measured, and the Tunisian transcripts were not normalized.
  • The split was made by file, so speakers do not overlap: Train has 13 h and 37 speakers, while Dev (1 h) and Test (50 min) each have 2 male speakers.
  • ASR test results:
    • Whisper Large-v3 improves from 76.46 WER zero-shot to 29.94 WER / 11.57 CER after fine-tuning.
    • W2v-Bert + CTC reaches 28.03 WER / 9.81 CER.
  • STT test BLEU:
    • The best cascade, Whisper Large-v3 → NLLB 3.3B, reaches 18.34.
    • The best end-to-end system is Whisper Large-v3 + NLLB 1.3B at 15.68 with pre-adapted components, versus 11.91 when starting from the original checkpoints.
    • Larger models helped consistently across these experiments.

Interesting ideas

  • I1. The transcription rules separate unadapted foreign words (Latin script) from loanwords that follow local pronunciation (Arabic script). They also attach native clitics to foreign stems and fix one spelling for numbers, days and percentages, which keeps mixed-language text consistent.
  • I2. Splitting by source file keeps speakers out of more than one partition, so the test set cannot reward memorized voices or styles.
  • I3. Each pretrained component is first adapted to the domain on its own. The components are then joined through a thin CNN bridge, and only the bridge and the layers next to it are trained.
  • I4. With scarce data, the cascade with an explicit intermediate transcript outperformed the end-to-end model by about 3 BLEU.
  • I5. Reporting zero-shot scores next to fine-tuned scores for every model size shows how much of the gain comes from the new data rather than from model scale.

For Mova

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

  • I1 (tokens): The tokenizer and lemmatizer could use a similar documented policy for borrowings, numerals and dates. One canonical form per item would keep downstream lemmas and measures stable.
  • I2 (learn): Silver data and treebanks could be split by document or author, so that evaluation and negative controls are not inflated by style leakage.
  • I3 (tree): gparse could be attached to a new encoder, or moved to a new UD dialect, by training only the interface layers. This keeps the rest frozen when little data is available.
  • I4 (frames): This supports keeping explicit, inspectable intermediate representations (tree, then frames) rather than collapsing stages, especially for low-resource inputs.
  • I5 (learn): Each evaluation could include an untrained or off-the-shelf baseline row, so that improvements from repair rules or silver data are clearly attributable.

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 · Machine translation · Natural language processing

Topics Speech & spoken language Semitic languages

 Institutions

🔬Laboratoire Informatique d'AvignonAvignon · France4 works · 414 citations🏢SystranParis · France2 works · 26 citations

Cite

Cite as Alex Choux, Marko Avila, Josep Crego, Fethi Bougares and Antoine Laurent. 2025. TuniFra: A Tunisian Arabic Speech Corpus with Orthographic Transcriptions and French Translations. Proceedings of The Third Arabic Natural Language Processing Conference.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.arabicnlp-main.5

@inproceedings{choux2025tunifra,
  title = {{TuniFra: A Tunisian Arabic Speech Corpus with Orthographic Transcriptions and French Translations}},
  author = {Alex Choux, Marko Avila, Josep Crego, Fethi Bougares and Antoine Laurent},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-main.5},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-main.5}
}

References

15 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ahmed Amine Ben Abdallah, Ata Kabboudi, Amir Kanoun, and Salah Zaiem. 2023. Leveraging data collection and unsupervised learning for code-switched tunisian arabic automatic speech recognition. Preprint, arXiv:2309.11327. ↗ our card
  2. Salima Mdhaffar, Fethi Bougares, Renato de Mori, Salah Zaiem, Mirco Ravanelli, and Yannick Estève. 2024. TARIC-SLU: A Tunisian benchmark dataset for spoken language understanding. In LREC-COLING 2024, pages 15606–15616, Torino, Italia. ELRA and ICCL. find
  3. Ahmed Ali, Peter Bell, James Glass, Yacine Messaoui, Hamdy Mubarak, Steve Renals, and Yifan Zhang. 2016. The mgb-2 challenge: Arabic multi-dialect broadcast media recognition. pages 279–284. our card
  4. 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. find
  5. Hedi Naouara, Jérôme Louradour, and Jean-Pierre Lorré. 2025. Linto audio and textual datasets to train and evaluate automatic speech recognition in tunisian arabic dialect. our card
  6. Good Data Workshop, AAAI 2025. find
  7. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine Mcleavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 28492–28518. PMLR. our card
  8. Ahmed Ali, Stephan Vogel, and Steve Renals. 2017. Speech recognition challenge in the wild: Arabic mgb-3. In 2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 316– 322. our card
  9. Bashar Talafha, Abdul Waheed, and Muhammad Abdul-Mageed. 2023. N-shot benchmarking of whisper on diverse arabic speech recognition. pages 5092–5096. our card
  10. 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. our card
  11. NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, and 20 others. 2022. No language left behind: Scaling human-centered machine translation. Preprint, arXiv:2207.04672. ↗
  12. Marko Avila and Josep Crego. 2025. Leveraging large pre-trained multilingual models for high-quality speech-to-text translation on industry scenarios. In Proceedings of the 31st International Conference on Computational Linguistics, pages 7624–7633, Abu Dhabi, UAE. Association for Computational Linguistics. find
  13. Abdul Waheed, Bashar Talafha, Peter Sullivan, Abdel-Rahim Elmadany, and Muhammad Abdul-Mageed. 2023. Voxarabica: A robust dialect-aware arabic speech recognition system. pages 441–449. find
  14. Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems, volume 33, pages 12449–12460. Curran Associates, Inc. our card
  15. Tianyi Xu, Kaixun Huang, Pengcheng Guo, Yu Zhou, Longtao Huang, Hui Xue, and Lei Xie. 2024. Towards rehearsal-free multilingual asr: A lora-based case study on whisper. pages 2534–2538. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/choux-2025-tunifra-tunisian-arabic-speech-corpus/