Elamur Research⌕
☰
home / papers / 2026 / WhiteHouse: Translation of the Casablanca Corpus for Multi-dialectal Arabic Speech Translation

WhiteHouse: Translation of the Casablanca Corpus for Multi-dialectal Arabic Speech Translation

Summary

Speech work on Arabic dialects has mostly targeted dialect identification and speech recognition. Speech translation from dialectal Arabic into English has had very little data, and the existing sets each cover only one dialect, such as Egyptian, Tunisian or North Levantine. The authors fill this gap with WhiteHouse. They took the publicly released development and test portions of Casablanca, a multi-dialectal ASR corpus built from TV-series episodes, and added a manual English translation to every transcribed utterance. The result is a three-way aligned resource: audio, dialectal transcript and English text, covering eight dialects. Translators followed the LDC Arabic-to-English guidelines in a two-pass workflow. A first pass drafted translations, with the audio available, and a second pass checked them against the guidelines. The authors then benchmarked current multilingual speech translation models on WhiteHouse, both zero-shot and after fine-tuning. These models performed weakly on dialectal input, which shows the resource is needed. Everything is openly published on Hugging Face as a shared benchmark for dialect speech translation.

Key points

  • WhiteHouse is presented as the first multi-dialectal Arabic→English speech translation corpus.
  • The eight dialects it spans are Moroccan, Egyptian, Yemeni, Algerian, Palestinian, Jordanian, Emirati and Mauritanian (Hassaniya).
  • Each dialect has about one hour of validation audio and one hour of test audio. That gives roughly 6.8k segments per split, with 667 to 1,045 segments per dialect.
  • The source is Casablanca (Talafha et al., 2024), a 48-hour TV-series corpus. Only its dev and test subsets are public, and these were also used in the NADI 2025 shared task.
  • The paper surveys related resources: Callhome Egyptian, ArzEnST, the UFAL North Levantine corpus, the LDC Tunisian telephone speech corpus used at IWSLT, and TEDxTN.

Interesting ideas

  • I1. Instead of collecting new audio, the authors added a translation layer to an existing transcribed corpus, which yields a parallel resource cheaply and keeps it aligned with prior ASR benchmarks.
  • I2. Annotation runs in two passes: a fast draft, then a separate check against a written guideline, which catches both inattention errors and guideline violations.
  • I3. Translators may consult the audio whenever the transcript is doubtful, so the translation does not inherit errors from the upstream transcription layer.
  • I4. Results are reported separately for each dialect and for both zero-shot and fine-tuned settings, so weaknesses hidden by a single average become visible.
  • I5. The paper argues that varieties without a fixed spelling make an intermediate written layer a weak point. Going end-to-end avoids relying on unstable surface forms.

For Mova

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

  • I1 (learn): Mova could add frame and role annotations on top of existing UD treebanks rather than new text, so parser and frame evaluations share one aligned test set.
  • I2 (learn): LLM teacher tiers could follow the same pattern, with one tier drafting silver labels and another auditing them against explicit annotation rules before training.
  • I3 (frames): when assigning semantic roles, the frame stage could look back to raw tokens and the original sentence when the UD tree looks suspicious, rather than trusting the tree alone.
  • I4 (tree): parser scores such as LAS could be broken down by genre and by the source UD treebank's dialect conventions, exposing where conversion or domain shift hurts.
  • I5 (tokens): this is a reminder to normalize noisy or nonstandard English spellings explicitly in the tokenizer and lemmatizer, so variant forms do not break the explainable pipeline downstream.

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

Topics Speech & spoken language Semitic languages

Cite

Cite as Fethi Bougares, Salima Mdhaffar and Yannick Estève. 2026. WhiteHouse: Translation of the Casablanca Corpus for Multi-dialectal Arabic Speech Translation. Proceedings of the Fifteenth Language Resources and Evaluation Conference.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/4zqn965acien

@inproceedings{bougares2026whitehouse,
  title = {{WhiteHouse: Translation of the Casablanca Corpus for Multi-dialectal Arabic Speech Translation}},
  author = {Fethi Bougares, Salima Mdhaffar and Yannick Estève},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/4zqn965acien},
  url = {https://doi.org/10.63317/4zqn965acien}
}

References

48 references, 14 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. 2024. Leveraging data collection and unsupervised learning for code-switched tunisian arabic automatic speech recognition. In ICASSP 2024. our card
  2. Idris Abdulmumin and al. 2025. Findings of the IWSLT 2025 evaluation campaign. In IWSLT 2025, pages 412–481, Vienna, Austria. ACL. find
  3. Milind Agarwal and al. 2023. find
  4. FINDINGS OF THE IWSLT 2023 EVALUATION CAMPAIGN. In IWSLT 2023, pages 1–61, Toronto. ACL. our card
  5. Ibrahim Said Ahmad and al. 2024. find
  6. FINDINGS OF THE IWSLT 2024 EVALUATION CAMPAIGN. In IWSLT 2024, pages 1–11, Bangkok. find
  7. Ahmed Ali, Suwon Shon, Younes Samih, Hamdy Mubarak, Ahmed Abdelali, James Glass, Steve Renals, and Khalid Choukri. 2020. The mgb-5 challenge: Recognition and dialect identification of dialectal arabic speech. In ASRU 2019. find
  8. Ahmed Ali, Stephan Vogel, and Steve Renals. 2017. Speech recognition challenge in the wild: Arabic mgb-3. our card
  9. Antonios Anastasopoulos and al. 2022. Findings of the IWSLT 2022 evaluation campaign. In Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 98–157, Dublin, Ireland. ACL. our card
  10. Jinze Bai and al. 2023. Qwen technical report. our card
  11. Alexandre Berard, Olivier Pietquin, Christophe Servan, and Laurent Besacier. 2016. Listen and translate: A proof of concept for end-to-end speech-to-text translation. find
  12. Fethi Bougares, Salima Mdhaffar, Haroun Elleuch, and Yannick Estève. 2025. TEDxTN: TEDx speech translation corpus for code-switched tunisian arabic - english. In WANLP 2025. find
  13. Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, Chang Zhou, and Jingren Zhou. 2024. Qwen2-audio technical report. arXiv preprint arXiv:2407.10759. ↗ our card
  14. Seamless Communication and al. 2023. Seam-lessm4t: Massively multilingual multimodal machine translation. find
  15. Alexis Conneau, Min Ma, Simran Khanuja, Yu Zhang, Vera Axelrod, Siddharth Dalmia, Jason Riesa, Clara Rivera, and Ankur Bapna. 2022. Fleurs: Few-shot learning evaluation of universal representations of speech. our card
  16. Mattia A. Di Gangi, Roldano Cattoni, Luisa Bentivogli, Matteo Negri, and Marco Turchi. 2019. MuST-C: a Multilingual Speech Translation Corpus. In NAACL. our card
  17. Amirbek Djanibekov, Hawau Olamide Toyin, Raghad Alshalan, Abdullah Alitr, and Hanan Aldarmaki. 2025. Dialectal coverage and generalization in arabic speech recognition. find
  18. Long Duong, Antonios Anastasopoulos, David Chiang, Steven Bird, and Trevor Cohn. An attentional model for speech translation without transcription. In NAACL 2016, San Diego. find
  19. Haroun Elleuch, Salima Mdhaffar, Yannick Estève, and Fethi Bougares. 2025. Adi-20: Arabic dialect identification dataset and models. In Interspeech, The Netherlands. our card
  20. Nizar Habash and al. 2015. Palestinian arabic conventional orthography guidelines - technical report. In Birzeit University and New York Univesity Abu Dhabi. find
  21. Nizar Habash and al. 2018. Unified guidelines and resources for Arabic dialect orthography. In LREC 2018, Miyazaki, Japan. our card
  22. 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. find
  23. Injy Hamed, Ngoc Thang Vu, and Slim Abdennadher. 2020. ArzEn: A speech corpus for code-switched Egyptian Arabic-English. In LREC 2020, pages 4237–4246, Marseille, France. ELRA. find
  24. Amir Hussein, Shinji Watanabe, and Ahmed Ali. 2021. Arabic speech recognition by end-to-end, modular systems and human. our card
  25. Ali Can Kocabiyikoglu, Laurent Besacier, and Olivier Kraif. 2018. Augmenting librispeech with french translations: A multimodal corpus for direct speech translation evaluation. find
  26. Ajinkya Kulkarni and Hanan Aldarmaki. 2023. Yet another model for arabic dialect identification. find
  27. Gaurav Kumar, Yuan Cao, Ryan Cotterell, Chris Callison-Burch, Daniel Povey, and Sanjeev Khudanpur. 2014. Translations of the callhome Egyptian Arabic corpus for conversational speech translation. In SLT 2014, pages 244–248, Lake Tahoe, California. find
  28. Linguistic Data Consortium LDC. 2013. Bolt program: Arabic to english translation guidelines. find
  29. Rao Ma, Mengjie Qian, Yassir Fathullah, Siyuan Tang, Mark Gales, and Kate Knill. 2025. Cross-lingual transfer learning for speech translation. find
  30. Abir Masmoudi, Fethi Bougares, Mariem Ellouze, Yannick Estève, and Lamia Belguith. 2018. Automatic speech recognition system for tunisian dialect. Lang. Resour. Eval., page 249–267. find
  31. Salima Mdhaffar, Fethi Bougares, Renato De Mori, Salah Zaiem, Mirco Ravanelli, and Yannick Estève. 2024a. Taric-slu: A tunisian benchmark dataset for spoken language understanding. In LREC-COLING 2024. find
  32. Salima Mdhaffar, Haroun Elleuch, Fethi Bougares, and Yannick Estève. 2024b. Performance analysis of speech encoders for low-resource slu and asr in tunisian dialect. In WANLP, pages 130– 139. find
  33. Matt Post. 2018. A call for clarity in reporting bleu scores. our card
  34. 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
  35. Mirco Ravanelli, Titouan Parcollet, Adel Moumen, Sylvain De Langen, Cem Subakan, Peter Plantinga, Yingzhi Wang, Pooneh Mousavi, Luca Della Libera, Artem Ploujnikov, et al. 2024. Open-source conversational ai with speechbrain 1.0. Journal of Machine Learning Research, 25(333):1–11. find
  36. Elizabeth Salesky, Matthias Sperber, and Alan W Black. 2019. Exploring phoneme-level speech representations for end-to-end speech translation. find
  37. 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. find
  38. Gregory Sanders, Sébastien Bronsart, Sherri Condon, and Craig Schlenoff. 2008. Odds of successful transfer of low-level concepts: a key metric for bidirectional speech-to-speech machine translation in DARPA’s TRANSTAC program. In LREC 2008, Marrakech, Morocco. find
  39. Suwon Shon, Ahmed M. Ali, Younes Samih, Hamdy Mubarak, and James R. Glass. 2020. Adi17: A fine-grained arabic dialect identification dataset. ICASSP 2020, pages 8244–8248. our card
  40. Tzu-Wei Sung, Jun-You Liu, Hung yi Lee, and Lin-Shan Lee. 2019. Towards end-to-end speech-to-text translation with two-pass decoding. ICASSP 2019, pages 7175–7179. find
  41. Bashar Talafha and al. 2025. Nadi 2025: The first multidialectal arabic speech processing shared task. find
  42. 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. ↗
  43. Changhan Wang, Juan Pino, Anne Wu, and Jiatao Gu. 2020a. CoVoST: A diverse multilingual speech-to-text translation corpus. In LREC, pages 4197–4203, Marseille, France. find
  44. Changhan Wang, Anne Wu, and Juan Pino. 2020b. Covost 2 and massively multilingual speech-to-text translation. find
  45. Petr Zemánek, Adam Pospíšil, Hashem Sellat, Mateusz Krubiński, and Pavel Pecina. 2023. UFAL speech corpus of north levantine arabic 1.0 - part 1. LINDAT/CLARIAH-CZ digital library at the Institute of Formal and Applied Linguistics (ÚFAL). find
  46. Biao Zhang, Ivan Titov, and Rico Sennrich. 2019. Improving deep transformer with depth-scaled initialization and merged attention. find
  47. Ruiqing Zhang, Xiyang Wang, Chuanqiang Zhang, Zhongjun He, Hua Wu, Zhi Li, Haifeng Wang, Ying Chen, and Qinfei Li. 2021. Bstc: A large-scale chinese-english speech translation dataset. find
  48. Inès Zribi, Rahma Boujelbane, Abir Masmoudi, Mariem Ellouze, Lamia Belguith, and Nizar Habash. 2014. A conventional orthography for Tunisian Arabic. In LREC 2014, pages 2355– 2361, Reykjavik, Iceland. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/bougares-2026-whitehouse-translation-casablanca-corpus-multi/