Metapseud at Iqra’Eval: Domain Adaptation with Multi-Stage Fine-Tuning for Phoneme-Level Qur’anic Mispronunciation Detection
Year: 2025
Type: conference-paper
Abstract: This paper presents the Metapseud system designed for the Iqra'Eval shared task, which addresses the automatic assessment of Qur'anic recitation pronunciation, as part of ARABIC-NLP 2025.This system applies multi-stage fine-tuning of Wav2Vec2.0with curriculuminspired training, followed by domain adaptation to Qur'anic phoneme annotations.The decoding is improved using beam search with a CTC-based decoder.The results show that staged adaptation achieved a phoneme error rate (PER) of 21% in the development set, and beam search improves the accuracy in the open test set from 76.9% to 82.1%.The findings of this work emphasize the significance of curriculum learning, domain adaptation, and decoding strategies in recognizing mispronunciation in Qur'anic recitation.
Author: Ayman Mansour
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037806