Our preprint, A Comparative Study of Pretrained Transformer Models for Quranic ASR: Speech Representations, Label Formats, and Dataset Composition, is now available online.
This work was led primarily by Nabil Mosharraf Hossain, with co-authors Riasat Islam and Unaizah Obaidellah. It is currently available as an arXiv preprint and is under review with a reputed journal. That means the work is public and readable, but it should still be treated as a preprint until peer review is complete.
You can read the publication page here: A Comparative Study of Pretrained Transformer Models for Quranic ASR.
Why Quranic ASR matters
Automatic speech recognition, or ASR, is the technology that converts spoken audio into text. In everyday settings, it powers voice assistants, dictation tools, captioning systems, and search over audio.
Quranic ASR is more specialised.
It focuses on recognising Qur’an recitation. This is a different problem from ordinary Arabic speech recognition because Qur’an recitation has distinctive pronunciation, rhythm, elongation, pauses, tajwid-related features, and variation across reciters and recording conditions.
A reliable Quranic ASR system could support many beneficial applications:
- searching within recitation audio
- aligning recitation with Qur’anic text
- building learning tools for students
- supporting memorisation and revision workflows
- improving accessibility for Islamic audio collections
- helping researchers analyse large recitation datasets
But because the Qur’an is sacred text, accuracy and care matter deeply. The goal is not only to produce technically impressive models. The goal is to understand what works, what fails, and what design decisions make recognition systems more reliable in this domain.
What the preprint studies
The paper compares pretrained Transformer models for Quranic automatic speech recognition.
Specifically, it studies models including Wav2Vec2.0, HuBERT, and XLS-R. These models learn useful representations from speech and can then be fine-tuned for recognition tasks.
The study looks at several practical choices that can affect performance:
- which pretrained speech representation is used
- how output labels are formatted
- whether the Arabic text includes diacritics
- how training strategies affect results
- how dataset composition changes model behaviour
This matters because performance in ASR is not determined by the model alone. Data quality, label design, preprocessing, and training setup can all change the final result.
The dataset: over 870 hours of recitation
The models were fine-tuned on a filtered Quranic dataset exceeding 870 hours of professional and user recitations.
That scale is important. Qur’an recitation varies across reciters, microphones, recording environments, and styles. A system trained on narrow or overly clean data may not generalise well to real-world use.
By comparing model behaviour across a large filtered dataset, the paper gives a more grounded view of what current Transformer-based ASR models can do for Quranic recitation.
What the results suggest
The best-performing configuration achieved lower word error rates than a Citrinet baseline while substantially reducing training time.
One important finding is that Arabic text without diacritics worked particularly well for this task. That is a useful practical result because label format is a major design choice in Arabic and Quranic ASR.
The paper also identifies Wav2Vec2-XLSR-53 as the strongest overall speech representation in this setting.
These findings do not mean the problem is solved. They help clarify which modelling choices appear most promising and where future research can build.
Why label format matters
Arabic presents special challenges for speech recognition. Diacritics carry important pronunciation and meaning information, but they also make the output space more complex.
For Quranic recitation, the question becomes even more sensitive. The written text, the recited sound, and the expected output representation are closely related but not identical design problems.
Choosing whether to model diacritics directly is not only a linguistic question. It is also a machine learning question. A more detailed label format may preserve more information, but it may also increase model difficulty and error rates.
The preprint’s finding that undiacritised Arabic text performed strongly gives developers and researchers a practical baseline for future Quranic ASR systems.
Why this connects to Islamic Computing
This work fits into the wider area of Islamic Computing: building computational systems that serve Islamic learning, access, preservation, and practice.
Quranic ASR is not simply another speech recognition benchmark. It sits at the intersection of AI, Arabic language processing, Qur’anic studies, recitation practice, and educational technology.
That means technical progress should be paired with domain responsibility.
Systems that handle Qur’an recitation should be evaluated carefully. They should communicate uncertainty. They should be designed with respect for the sacred nature of the content. And when used in learning contexts, they should support students rather than replace qualified teachers or careful human review.
What builders can take from the paper
For researchers and developers working on Islamic technology, the paper suggests several lessons.
First, start with strong baselines. Comparing multiple Transformer representations helps clarify whether improvements come from the model, the data, or the training setup.
Second, treat label design as a serious research decision. In Arabic and Quranic ASR, label format can shape performance significantly.
Third, care about dataset composition. Recitation audio is diverse, and systems should be evaluated against that diversity.
Fourth, do not separate performance from responsibility. A lower word error rate is useful, but applications involving Qur’an recitation also need trust, careful presentation, and appropriate human oversight.
Fifth, recognise the value of focused student-led research. This project is a good example of how meaningful Islamic technology research can emerge when students combine technical depth with a problem that genuinely matters.
In summary
The preprint shows that pretrained Transformer models can be effective for Quranic ASR, especially when model choice, label format, and dataset composition are treated carefully.
Nabil’s work helps move this area forward by giving a comparative view of what current speech models can do for Qur’an recitation and where future systems can improve.
Read the preprint details here: A Comparative Study of Pretrained Transformer Models for Quranic ASR.
The arXiv record is also available directly at https://arxiv.org/abs/2606.19747.
Leave a Reply