Using LALMs for pronunciation error detection in Arabic

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This paper presents Qwen-MDD-AR, our system for the Iqra’Eval.2 Interspeech 2026 challenge on Modern Standard Arabic pronunciation assessment. Built on Qwen3-ASR-1.7B, it frames Arabic mispronunciation detection and diagnosis as autoregressive speech-to-phoneme generation. We add the 68-phoneme MSA inventory as special tokens, freeze the audio encoder, and adapt the decoder with LoRA. Using only provided challenge data, our submission ranked 6th of 19 teams on the hidden QuranMB.v2 test set, with 0.6702 F1 and 0.0445 PER. In a controlled comparison on our internal mixed held-out set, a Wav2Vec2-BERT CTC baseline trained on the same data achieved better PER, precision, and F1, while Qwen-MDD-AR obtained the same recall and a small Correct Diagnosis advantage (0.9496 vs. 0.9328). Results suggest LoRA-adapted LALMs are a practical alternative for Arabic MDD, while alignment accuracy and inference cost remain open.