IqraEval.2 Challenge -- Interspeech 2026

[Yassine El Kheir]

Dear colleagues,

We are pleased to announce that IqraEval V2 is now publicly available as part of the IqraEval.2 Challenge (Interspeech 2026), focusing on automatic mispronunciation detection in Modern Standard Arabic (MSA).

The challenge targets phoneme-level detection and diagnosis of pronunciation errors (substitution, deletion, insertion) from read, fully diacritized MSA speech.

- What’s new in IqraEval V2?

In addition to native and synthetic data, IqraEval V2 includes a newly released dataset of real human mispronunciations, collected specifically for Interspeech 2026.
This dataset can be used as training or development data and complements the existing resources.

- Available Public Datasets

All data are fully public and hosted on Hugging Face:

• Native MSA speech (pseudo-labeled)

• Synthetic mispronunciations (TTS-based)

• Real mispronunciations (human-recorded, Interspeech 2026)

• Evaluation set

Strict checks ensure no speaker or sentence overlap between training and evaluation data.

👉 Full task description, datasets, phoneme set, and evaluation protocol:
https://huggingface.co/spaces/IqraEval/IqraEval_Interspeech_26

🤖 Models and Tools

• All baseline models and utilities are publicly available:
  https://huggingface.co/IqraEval

• Includes an open-source MSA phonetizer (requires fully diacritized input).

🏆 Leaderboard & Submission

• The leaderboard and submission details will be released soon.

• The primary ranking metric will be phoneme-level F1-score.

We warmly invite the community to explore the data, experiment with models, and prepare for participation in the upcoming challenge.

If you have any questions or are interested in participating, please don't hesitate to reach out.

Best regards,
Yassin El Kheir, 
On Behalf of IqraEval Team