A Comparative Study of Pretrained Transformer Models for Quranic ASR: Speech Representations, Label Formats, and Dataset Composition

Two recent studies on arXiv explore the application of advanced AI models in distinct domains: Quranic Automatic Speech Recognition (ASR) and on-device fault detection using lightweight transformer architectures.

A comparative study on Quranic ASR fine-tunes pretrained Transformer-based models to improve transcription accuracy, achieving a Word Error Rate (WER) of 0.08 on the EveryAyah subset and 0.11 on the combined EveryAyah+Tarteel setting^[1]. The research identifies key factors affecting transcription accuracy and reduces combined-model training time from 140 hours to 40 hours. Meanwhile, a benchmark study compares traditional machine learning methods with lightweight transformer architectures for on-device fault detection, finding that TinyBERT-4L is the most deployment-friendly transformer at 55 MB and 18 ms CPU latency^[2]. The study also shows that INT8 quantization reduces model size by 25% while preserving 86.9% F1-score, and an adaptive inference pipeline achieves 87.6% F1-score at 19.5 ms average latency. These advancements have implications for applications such as aided memorisation tools, Quranic search engines, and resource-constrained hardware.