New AI System Detects and Corrects Pronunciation Errors in Quranic Recitation

A team of researchers has introduced an automated system capable of identifying and correcting mispronunciations in Quranic recitation by leveraging Islamic recitation principles and machine learning. The work represents a significant advance in applying artificial intelligence to religious education, addressing both technical challenges in speech assessment and cultural requirements unique to Quranic study.

The Challenge of Scale and Quality#

Evaluating spoken language through machine learning typically requires enormous quantities of carefully labeled training data—a particular problem for specialized domains like Quranic recitation. The researchers overcame this barrier by creating an automated pipeline that produces high-quality datasets at scale. Their approach collected recordings from expert reciters, segmented audio according to traditional pause points, and verified transcripts using a novel verification algorithm they call Tasmeea. The result: 848 hours of annotated audio containing over 286,000 utterances ready for model training.

Building a Custom Framework#

Rather than applying standard speech recognition tools designed for general Arabic or other languages, the team developed a custom phonetic encoding system tailored to Quranic requirements. Called the Quran Phonetic Script (QPS), this system encodes both the individual phonemes and the articulation characteristics specific to proper Quranic recitation. At its core, QPS operates on two levels: phoneme encoding (capturing Arabic letters and vowel length) and articulation features (capturing how each sound should be produced). The researchers also created a benchmark dataset called qdat_bench to evaluate performance on real recitation errors, covering phonetic, diacritization, and Tajweed rules (Islamic recitation principles governing sounds like Ghunnah, Qalqalah, and Madd).

Achieving High Accuracy#

The team trained a multi-level deep learning model and achieved remarkable precision: a phoneme error rate of just 0.21% on their test set and 1.94% on the qdat_bench benchmark. The system also scored a 75.8% F1 score when detecting Tajweed rule violations. These results suggest the technology could effectively support learners of Quranic recitation by providing immediate, rule-based feedback.

Impact and Availability#

The researchers have released their work as open-source, making the datasets, models, and code publicly available to the broader research community. This decision democratizes access to advanced Quranic speech processing technology and may inspire similar efforts in other religious and cultural education domains.