UBC-NLP/NADI2025_subtask2_ASR

Dataset Summary

This datasets is related to NADI-2015 Subtask-2: Multidialectal Arabic ASR.

This subtask aims to develop ASR systems that transcribe Arabic speech across dialects. This subtask is hosted on Codabench. Participants must handle phonetic and dialectal variation using the provided training and validation data. A private test set will be released soon. Models can be evaluated zero-shot on validation data or fine-tuned using the training set. For the SER subtask, participants will evaluate directly on the test set without access to training or validation data.

---

Access

Please register for the shared task.

• Register here: https://forms.gle/WHsyFMtyaewufN7E8
• Codabench → https://www.codabench.org/competitions/8587
• Website → https://nadi.dlnlp.ai/2025/

---

Contact & Links

• 🌐 Website: https://nadi.dlnlp.ai/2025/
• 📝 Registration Form: https://forms.gle/WHsyFMtyaewufN7E8
• 📣 Google Group: https://groups.google.com/u/4/g/nadi-shared-task-2025
• 📩 Email: [email protected]

---

Citation

Kindly cite the following paper when using this dataset:

@article{talafha2024casablanca,
  title={Casablanca: Data and Models for Multidialectal Arabic Speech Recognition},
  author={Talafha, Bashar and Kadaoui, Karima and Magdy, Samar Mohamed and Habiboullah, Mariem and Chafei, Chafei Mohamed and El-Shangiti, Ahmed Oumar and Zayed, Hiba and Alhamouri, Rahaf and Assi, Rwaa and Alraeesi, Aisha and others},
  journal={arXiv preprint arXiv:2410.04527},
  year={2024}
}

License: BigScience OpenRAIL-M

This dataset is released under the BigScience OpenRAIL-M license and is intended for academic research and non-commercial use only. Commercial use, redistribution, or integration into products is prohibited without prior written permission.