Elamur Research⌕
☰
home / papers / 2025 / TEDxTN: A Three-way Speech Translation Corpus for Code-Switched Tunisian Arabic - English

TEDxTN: A Three-way Speech Translation Corpus for Code-Switched Tunisian Arabic - English

Paper · 2025 · Fethi Bougares, Salima Mdhaffar, Haroun Elleuch et al.

Summary

Tunisian Arabic has little public speech data, and almost none that pairs code-switched speech with translations. Tunisian speakers mix their dialect with French and, more and more, English. The authors built TEDxTN to fill this gap. It is an openly released three-way resource that links Tunisian audio, its transcription and an English translation. They took 108 TEDx talks, about 25 hours of speech from speakers in 11 regions. Professional annotators segmented, transcribed and translated the talks under in-house guidelines, which are also published. The authors then trained baselines by fine-tuning pre-trained models in SpeechBrain. For recognition, the best model was w2v-Bert-2.0T, reaching 21.37% WER on test. For speech translation, fine-tuned Whisper-large-v3 reached 25.68 BLEU, compared with about 14 BLEU zero-shot. An error breakdown showed that many apparent recognition mistakes on mixed segments are correct words written in a different script than the reference. The resource matters because it is the first public benchmark of its kind for Tunisian code-switched speech translation.

Key points

  • TEDxTN contents: 108 TEDx talks, roughly 25 hours of code-switched Tunisian speech, from speakers across more than 11 Tunisian regions.
  • Release and license: The data, guidelines, models and a SpeechBrain recipe are on Hugging Face (fbougares/TedxTn). Because of the CC BY-NC-ND 4.0 license, only transcripts and translations are distributed. The audio is referenced by YouTube URLs.
  • ASR results: w2v-Bert-2.0T gets 19.92% WER on dev and 21.37% on test. By segment type, test WER is 24.16% for Tunisian-only, 21.31% for mixed and 13.93% for fully foreign segments.
  • Speech translation results: Whisper models were fine-tuned for 80 epochs. Large-v3 improves from 10.96/13.95 BLEU zero-shot to 25.19/25.68 (valid/test). Scores use truecased BLEU without punctuation via sacrebleu.
  • Failed variant: Feeding the WLSR or w2v-Bert-2.0T encoders into an NLLB 1.3B decoder stayed below 5 BLEU. The authors leave this for future work.

Interesting ideas

  • I1. The authors split the test set by language makeup (dialect-only, mixed, foreign-only) and report error per slice, which shows where a model actually struggles instead of hiding it in one average.
  • I2. Manual inspection found that many "errors" were correct words written in the other script, so the metric was penalising spelling conventions rather than recognition.
  • I3. The paper names a typology of switching: between sentences, inside a sentence, tag-like insertions, and intra-word mixing where native affixes attach to foreign stems.
  • I4. To respect a no-derivatives license, the team ships only its own annotations plus pointers to the original media, not the source itself.
  • I5. Keeping audio, transcript and translation aligned lets the same data test both direct translation and a pipeline through an intermediate representation.

For Mova

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

  • I1 (learn): Evaluation could report parser and reader scores separately for slices such as plain text, text with foreign spans, and dialogue, so a regression in one slice cannot hide behind a stable overall score.
  • I2 (tokens): Before scoring tokenizer or lemma output, normalise known variant spellings (British/American, hyphenation, transliterated names) so that convention mismatches are counted separately from real mistakes.
  • I3 (tree): The same categories could guide how foreign spans in English text are tagged and attached in UD (e.g., flat:foreign versus a normal head), with a separate rule for borrowed stems carrying English affixes.
  • I4 (learn): Silver corpora built from restricted texts could be stored as document IDs, offsets and annotations, with the raw text fetched locally, which keeps the shared library legally clean.
  • I5 (back): A triple of source text, meaning graph and reference translation would allow comparing translation through the meta-world against a direct baseline on identical items, showing where the graph loses information.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🔬Laboratoire Informatique d'AvignonAvignon · France7 works · 436 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSignal Processing · 452 works

🧭 Topics 3Research subjects and languages the work is about.English · 2,263 worksSemitic languages · 2,908 worksSpeech & spoken language · 480 works

Cite

Cite as Fethi Bougares, Salima Mdhaffar, Haroun Elleuch and Yannick Estève. 2025. TEDxTN: A Three-way Speech Translation Corpus for Code-Switched Tunisian Arabic - English. 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.22

@inproceedings{bougares2025tedxtn,
  title = {{TEDxTN: A Three-way Speech Translation Corpus for Code-Switched Tunisian Arabic - English}},
  author = {Fethi Bougares, Salima Mdhaffar, Haroun Elleuch and Yannick Estève},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-main.22},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-main.22}
}

References

48 references, 16 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. The Cambridge Handbook of Linguistic Code-switching. Cambridge Handbooks in Language and Linguistics. Cambridge University Press, 2009. find
  2. 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. ↗
  3. Ghazi Abuhakema. 2013. Code switching and code mixing in arabic written advertisements: Patterns, aspects, and the question of prestige and standardisation. find
  4. Kais Sultan Mousa Alowidha. 2024. English-arabic code switching and identity in bilingual saudis living in saudi arabia: A comparative study between large and small cities. Educational Administration: Theory and Practice, 30(5). find
  5. Antonios Anastasopoulos, Loïc Barrault, Luisa Bentivogli, Marcely Zanon Boito, Ondřej Bojar, Roldano Cattoni, Anna Currey, Georgiana Dinu, Kevin Duh, Maha Elbayad, Clara Emmanuel, Yannick Estève, Marcello Federico, Christian Federmann, Souhir Gahbiche, Hongyu Gong, Roman Grundkiewicz, Barry Haddow, Benjamin Hsu, and 24 others. 2022. Findings of the IWSLT 2022 evaluation campaign. In IWSLT, pages 98–157, Dublin, Ireland.
  6. Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, and Michael Auli. 2021. Xls-r: Self-supervised cross-lingual speech representation learning at scale. Preprint, arXiv:2111.09296. ↗
  7. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2016. Neural machine translation by jointly learning to align and translate. Preprint, arXiv:1409.0473. ↗
  8. Marcely Zanon Boito, John Ortega, Hugo Riguidel, Antoine Laurent, Loïc Barrault, Fethi Bougares, Firas Chaabani, Ha Nguyen, Florentin Barbier, Souhir Gahbiche, and 1 others. 2022. On-trac consortium systems for the iwslt 2022 dialect and low-resource speech translation tasks. find
  9. Chung-Pu Chiou, Hou-An Lin, and Chia-Ping Chen. 2022. Mandarin-English code-switching speech recognition system for specific domain. In RO-CLING. find
  10. Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder-decoder for statistical machine translation. Preprint, arXiv:1406.1078. ↗
  11. S. A. Chowdhury, A. Hussein, Ahmed Abdelali, and Ahmed Ali. 2021. Towards one model to rule all: Multilingual strategy for dialectal code-switching arabic asr. ArXiv, abs/2105.14779. ↗
  12. Seamless Communication, Loïc Barrault, Yu-An Chung, Mariano Coria Meglioli, David Dale, Ning Dong, Mark Duppenthaler, Paul-Ambroise Duquenne, Brian Ellis, Hady Elsahar, Justin Haaheim, and all. 2023. Seamless: Multilingual expressive and streaming speech translation. Preprint, arXiv:2312.05187. ↗
  13. Amitava Das and Björn Gambäck. 2014. Identifying languages at the word level in code-mixed Indian social media text. In ICON, Goa, India. find
  14. Anik Dey and Pascale Fung. 2014. A Hindi-English code-switching corpus. In LREC. find
  15. Ahmad Abdel Tawwab Sharaf Eldin. 2014. Socio linguistic study of code switching of the arabic language speakers on social networking. International Journal of English Linguistics, 4:78. ↗
  16. Mohammed O. Elfahal, Mohammed Elhafiz Mustafa Supervisor, and Rashid A. Saeed Co-Supervisor. 2019. Automatic recognition and identification for mixed sudanese arabic – english languages speech. find
  17. Alex Graves, Abdel rahman Mohamed, and Geoffrey E. Hinton. 2013. Speech recognition with deep recurrent neural networks. ICASSP. find
  18. Nizar Habash, Fadhl Eryani, Salam Khalifa, Owen Rambow, Dana Abdulrahim, Alexander Erdmann, Reem Faraj, Wajdi Zaghouani, Houda Bouamor, Nasser Zalmout, Sara Hassan, Faisal Al-Shargi, Sakhar B. Alkhereyfy, Basma Abdulkareem, Ramy Eskander, Mohammad Salameh, and Hind Saddiki. 2018. Unified guidelines and resources for arabic dialect orthography. In LREC.
  19. Injy Hamed, Mohamed Elmahdy, and Slim Abdennadher. 2018. Collection and analysis of code-switch Egyptian Arabic-English speech corpus. In LREC. find
  20. Injy Hamed, Nizar Habash, Slim Abdennadher, and Ngoc Thang Vu. 2022. ArzEn-ST: A three-way speech translation corpus for code-switched Egyptian Arabic-English. In WANLP. ↗
  21. Injy Hamed, Caroline Sabty, Slim Abdennadher, Ngoc Thang Vu, Thamar Solorio, and Nizar Habash. 2025. A survey of code-switched Arabic NLP: Progress, challenges, and future directions. In COL-ING. ↗
  22. Injy Hamed, Ngoc Thang Vu, and Slim Abdennadher. 2020. ArzEn: A speech corpus for code-switched Egyptian Arabic-English. In LREC. find
  23. François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia Tomashenko, and Yannick Estève. 2018. TED-LIUM 3: Twice as Much Data and Corpus Repartition for Experiments on Speaker Adaptation, page 198–208. Springer International Publishing. find
  24. Linguistic Data Consortium LDC. 2013. Bolt program: Arabic to english translation guidelines. find
  25. Mirco Ravanelli, Titouan Parcollet, Adel Moumen, Sylvain de Langen, Cem Subakan, Peter Plantinga, Yingzhi Wang, Pooneh Mousavi, Luca Della Libera, Artem Ploujnikov, and 1 others. 2024. Open-source conversational ai with speechbrain 1.0. Journal of Machine Learning Research, 25(333):1–11. find
  26. Elizabeth Salesky, Matthew Wiesner, Jacob Bremerman, Roldano Cattoni, Matteo Negri, Marco Turchi, Douglas W. Oard, and Matt Post. 2021. The multilingual tedx corpus for speech recognition and translation. Preprint, arXiv:2102.01757. ↗
  27. Fung P. Li, Y. 2013. Language modeling for mixed language speech recognition using weighted phrase extraction. In INTERSPEECH 2013, Lyon, France. find
  28. Silin Li, Ruoyu Song, Tianwei Lan, Zeming Liu, and Yuhang Guo. 2024. TED-EL: A corpus for speech entity linking. In LREC-COLING 2024, pages 15721– 15731, Torino, Italia. ELRA and ICCL. find
  29. Ying Li and Pascale Fung. 2013. Improved mixed language speech recognition using asymmetric acoustic model and language model with code-switch inversion constraints. In ICASSP. find
  30. 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. find
  31. Lotfi Sayahi. 2011. change in tunisia. Code-switching and language Nivedita Sethiya and Chandresh Kumar Maurya. 2025. End-to-end speech-to-text translation: A survey. Comput. Speech Lang., 90(C). find
  32. Ganji Sreeram, Kunal Dhawan, and Rohit Sinha. 2018. Hindi-english code-switching speech corpus. CoRR, abs/1810.00662. ↗
  33. Kyoko Takashi. 1990. A sociolinguistic analysis of english borrowings in japanese advertising texts. World Englishes. Carol. ↗
  34. Myers-Scotton. 2013. Contact linguistics: Bilingual encounters and grammatical outcomes. Bashar Talafha, Karima Kadaoui, Samar Mohamed Magdy, Mariem Habiboullah, Chafei Mohamed Chafei, Ahmed Oumar El-Shangiti, Hiba Zayed, Mohamedou Cheikh Tourad, Rahaf Alhamouri, Rwaa Assi, Aisha Alraeesi, Hour Mohamed, Fakhraddin Alwajih, Abdelrahman Mohamed, and booktitle = El Mekki, Abdellah et al. Casablanca: Data and models for multidialectal Arabic speech recognition. find
  35. 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. ↗
  36. Good Data Workshop, AAAI 2025. find
  37. Bashar Talafha, Abdul Waheed, and Muhammad Abdul-Mageed. 2023. N-shot benchmarking of whisper on diverse arabic speech recognition. Preprint, arXiv:2306.02902. ↗
  38. Shana Poplack. 1980. Sometimes i’ll start a sentence in spanish y termino en espaÑol: toward a typology of code-switching 1. Linguistics, 18:581–618. find
  39. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. Attention is all you need. Preprint, arXiv:1706.03762. ↗
  40. Carol Myers-Scotton. 1989. Codeswitching with english: types of switching, types of communities. World Englishes, 8:333–346. ↗
  41. Matt Post. 2018. A call for clarity in reporting BLEU scores. In Conference on Machine Translation, pages 186–191, Brussels, Belgium. ACL.
  42. Shinji Watanabe, Takaaki Hori, Shigeki Karita, Tomoki Hayashi, Jiro Nishitoba, Yuya Unno, Nelson Enrique Yalta Soplin, Jahn Heymann, Matthew Wiesner, Nanxin Chen, Adithya Renduchintala, and Tsubasa Ochiai. 2018. Espnet: End-to-end speech processing toolkit. In Interspeech. find
  43. Daniel Povey, Arnab Ghoshal, Gilles Boulianne, Luká Burget, Ondrej Glembek, Nagendra Kumar Goel, Mirko Hannemann, Petr Motlícek, Yanmin Qian, Petr Schwarz, Jan Silovský, Georg Stemmer, and Karel Veselý. 2011. The kaldi speech recognition toolkit. find
  44. Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, and Michael Auli. 2023. Scaling speech technology to 1,000+ languages. Preprint, arXiv:2305.13516. ↗
  45. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust speech recognition via large-scale weak supervision. Preprint, arXiv:2212.04356. ↗
  46. Brian Yan, Patrick Fernandes, Siddharth Dalmia, Jiatong Shi, Yifan Peng, Dan Berrebbi, Xinyi Wang, Graham Neubig, and Shinji Watanabe. 2022. Cmu’s iwslt 2022 dialect speech translation system. In IWSLT 2022, pages 298–307. find
  47. Jinyi Yang, Amir Hussein, Matthew Wiesner, and Sanjeev Khudanpur. 2022. Jhu iwslt 2022 dialect speech translation system description. In IWSLT, pages 319– 326. find
  48. Chadi Ben Youssef and Stefan Th. Gries. 2023. Code-switching in tunisian arabic: A multifactorial random forest analysis. In Corpora, Volume 18 Issue 3. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/bougares-2025-tedxtn-three-way-speech-translation/