Tadabur: A Large-Scale Quran Audio Dataset
| Authors | Faisal Alherran |
| Year | 2026 |
| HF Upvotes | 5 |
| arXiv | 2604.18932 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
Despite growing interest in Quranic data research, existing Quran datasets remain limited in both scale and diversity. To address this gap, we present Tadabur, a large-scale Quran audio dataset. Tadabur comprises more than 1400+ hours of recitation audio from over 600 distinct reciters, providing substantial variation in recitation styles, vocal characteristics, and recording conditions. This diversity makes Tadabur a comprehensive and representative resource for Quranic speech research and analysis. By significantly expanding both the total duration and variability of available Quran data, Tadabur aims to support future research and facilitate the development of standardized Quranic speech benchmarks.
Engineering Breakdown
Plain English
Tadabur is a new large-scale audio dataset focused specifically on Quranic recitations, containing over 1400 hours of speech from more than 600 different reciters. The problem it solves is that existing Quran datasets are too small and lack diversity in speaker styles, vocal characteristics, and recording conditions—which limits training robust speech models for religious and linguistic analysis. The dataset's key value is its scale and variability: by capturing multiple reciters with different recitation styles and recording qualities, it provides the foundation for training generalized speech recognition and speaker identification models on Quranic content. This addresses a gap in the ML community where culturally and linguistically important datasets often lag behind general-purpose datasets.
Core Technical Contribution
The core contribution is the dataset itself rather than a novel algorithm: Tadabur provides an order-of-magnitude increase in both duration (1400+ hours) and speaker diversity (600+ reciters) for Quranic audio compared to prior public datasets. The novelty lies in systematic curation and aggregation of Quranic recitations while managing the practical challenges of diverse recording conditions and recitation styles—something that requires careful quality control and metadata annotation. By standardizing a benchmark specifically for Quranic speech, the authors enable reproducible research in an underserved domain and lay groundwork for developing Quranic speech processing benchmarks that the field currently lacks. The contribution is infrastructure-level: making high-quality, diverse data available unlocks downstream research that was previously bottlenecked by data scarcity.
How It Works
Tadabur aggregates Quranic recitation audio from multiple sources, likely including public repositories, religious institutions, and online platforms where Quranic recitations are distributed. For each audio file, the dataset captures or infers metadata including the reciter's identity, recitation style (e.g., Tajweed rules applied), and recording characteristics (sample rate, microphone type, acoustic environment). The input is raw audio in various formats and conditions; the transformation involves standardization (normalizing audio format, sample rates, and volume levels) and annotation (labeling reciter, recitation variant, and quality metrics). The output is a structured dataset with aligned audio files, reciter IDs, and metadata that researchers can use to train speech models, evaluate speaker identification, study prosody and recitation style variation, or benchmark Quranic speech processing systems. Quality control likely involves manual review and automated checks to ensure audio clarity and correct reciter attribution, addressing the challenge that diverse crowdsourced or archival recordings often have inconsistent quality.
Production Impact
For engineers building Quranic speech applications—such as recitation assistants, speaker identification systems, or speech-to-text for Islamic content—Tadabur immediately addresses data scarcity, reducing the need for expensive custom data collection. In a production pipeline, teams can now finetune pre-trained speech models (Whisper, Wav2Vec2, etc.) directly on this domain-specific data rather than struggling with tiny in-house datasets, dramatically improving accuracy and reducing time-to-deployment. The diversity of 600+ reciters means models trained on Tadabur generalize better to unseen speakers and recording conditions, reducing overfitting issues that plague small, homogeneous datasets. Trade-offs include the need to handle metadata alignment (mapping audio segments to correct speakers and Quranic chapters), potential licensing considerations for religious content, and the engineering overhead of streaming 1400+ hours of audio in training pipelines. The dataset also enables new product capabilities: speaker diarization, recitation style classification, and quality assessment tools that previously lacked sufficient labeled data.
Limitations and When Not to Use This
The paper does not present fine-grained evaluations showing how models trained on Tadabur perform on downstream tasks—no benchmarks are provided comparing accuracy on speaker identification, speech recognition, or other metrics against existing datasets or baselines. It's unclear whether the 1400 hours are balanced across reciters (some may have 10+ hours, others minimal), which would affect generalization; class imbalance in speaker distribution could bias models toward frequent reciters. The dataset focuses solely on Quranic recitation in Arabic, limiting applicability to research on other languages, religious traditions, or non-recitation speech domains; it also doesn't address how recording quality varies (some recitations may be archival low-fidelity, others modern high-quality), which could create confounds in downstream analyses. Finally, the paper lacks details on licensing, commercial use restrictions, and governance—critical for production systems that need clear data rights—and provides no discussion of privacy considerations for living reciters or how the dataset handles intellectual property of recorded recitations.
Research Context
This work builds on the broader trend of creating domain-specific, multilingual speech datasets (similar to efforts like Common Voice, VoxPopuli, or language-specific resources) and applies that infrastructure-building approach to an underrepresented cultural and linguistic domain. It directly addresses gaps identified in the speech processing community where Arabic speech datasets exist (e.g., AraBERT, Nuanced Arabic Speech) but Quranic-specific resources remain fragmented and limited, making it difficult to study properties unique to Quranic recitation like Tajweed rules and melodic variation. The dataset enables future research into speech phenomena specific to Islamic tradition: measuring how recitation style affects intelligibility, training speaker-independent Quranic recognition systems, or analyzing prosodic features across different recitation schools. This also opens the door to work at the intersection of cultural heritage preservation and AI—applying modern speech technology to digitize and make searchable massive archives of historical and contemporary Quranic recitations.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::