qArI: A Hybrid CTC/Attention-Based Model for Quran Recitation Recognition Using Bidirectional LSTMP in an End-to-End Architecture

The accurate speech recognition of the Holy Quran is crucial for maintaining the traditional recitation styles and pronunciations, which helps in preserving the authenticity of the Quranic teachings and ensuring their accurate transmission across generations. Though the application of freshly developed models to spoken and written Arabic and non-Arabic speech recognition has yielded highly accurate results, research on Holy Quran is still in its early levels. Indeed, speech recognition of the Holy Quran presents several challenges, including language complexity and the absence of a comprehensive dataset. This research aims to improve the accuracy of speech recognition models for the recital of the Holy Quran. A new dataset called comprehensive Quranic dataset version 1 (CQDV1) is created to serves the HQSR field. The dataset is publicly available for use by other researchers and includes recitations of the entire Quran (114 sura, recited by 35 reciters), based on Hafs from Asim narrative.The study explores the development of a speech recognition model for the accurate recital of the Holy Quran. The model combines a connectionist temporal classification (CTC)/attention loss function with a Bidirectional Long Short-Term Memory with projections (BLSTMP) architecture and a token-based recurrent neural network language model (RNNLM) using CQDV1 dataset. The results achieved were a token error rate (TER) of 6.4%, a word error rate (WER) of 10.4%, and a sentence error rate (SER) of 55.3% with
$\lambda =0.2$
.