New two-stage framework cracks Arabic mispronunciation detection with 63% improvement

Researchers have developed a new two-stage framework that significantly improves mispronunciation detection and diagnosis (MDD) for Modern Standard Arabic (MSA), a low-resource language where accurate phoneme recognition is especially challenging due to limited data and the gap between synthetic training data and real learner recordings. The system, detailed in a paper accepted to Interspeech 2026 and topping the IqraEval.2 Challenge, integrates a pre-trained encoder with causal dilated temporal convolutional networks to preserve fine-grained phonetic variations. It then uses a hierarchical two-stage strategy: first learning general mappings from native and synthetic corpora, then adapting to scarce real learner data to mitigate domain shift without over-correction. Prediction stability is further enhanced via multi-checkpoint ensemble inference with N-gram rescoring.

On the QuranMB.v2 test set, the framework achieved an F1-score of 0.7201, representing a 63.1% relative improvement over the baseline of 0.4414. This performance establishes a new state-of-the-art for low-resource MSA in MDD tasks. The approach is particularly notable for its effective handling of the synthetic-real domain gap, a common problem in speech tasks where models trained on clean or synthetic data fail to generalize to real-world, noisy learner audio. By combining pre-trained representations, temporal convolutional networks, and a two-stage adaptation scheme, the system offers a blueprint for improving pronunciation tools in other low-resource languages as well. The work was led by Jing Yang and seven co-authors from multiple institutions.