This benchmark contains Quranic recitation audio from three sources, including REAL USER PHONE RECORDINGS (Tarteel tlog) and studio recitations, redistributed here only to enable reproducible ASR evaluation. By requesting access you agree: (1) to use it solely for ASR research / evaluation, (2) NOT to redistribute the audio or use it for voice cloning / TTS / re-identification, and (3) to respect the reciters' and users' rights and the upstream dataset terms (tarteel-ai/tlog, tarteel-ai/everyayah, QUL).

Quranic ASR Benchmark - leakage-free, held-out

A small, leakage-free benchmark (600 clips) for evaluating Arabic ASR on Quranic recitation (Hafs riwayah). Every clip is verified absent from our training data, so it measures generalization, not memorization. Same clips + same scoring for every model.

📊 Live leaderboard: https://huggingface.co/spaces/Muno459/quranic-asr-leaderboard

The set (600 clips, 200 per source)

Source	n	What it is
`everyayah_heldout`	200	3 EveryAyah reciters with zero training samples (clean studio).
`qul_alnufais`	200	A QUL reciter (Al-Nufais) never trained on; references from the canonical mushaf.
`tlog_holdout`	200	Real phone-mic recitation held out from training (the hard, real-world case).

tlog is the one that matters for apps; studio numbers flatter every model.

Hugging Face loading

Each source is exposed as its own dataset configuration with a single test split. Rows include audio, text, id, and source, so ASR evaluation can be run without extra reference-transcript preprocessing.

from datasets import load_dataset

ds = load_dataset("Muno459/quranic-asr-benchmark", "tlog_holdout", split="test")
print(ds[0]["audio"], ds[0]["text"])

Available configs:

everyayah_heldout - clean studio held-out EveryAyah reciters
qul_alnufais - held-out QUL Al-Nufais reciter
tlog_holdout - held-out real phone-mic Tarteel tlog recordings

Files

benchmark.jsonl - one row per clip: id, source, reference_text, audio (relative path).
audio/<source>/*.wav - 16 kHz mono audio.
score.py - the official scorer (exact normalization + WER/CER, per-source + overall).
results.json - current leaderboard results.

How to evaluate your model

# 1. transcribe every clip in benchmark.jsonl (use the `audio` path), write one hyp per `id`:
#    {"id": "everyayah_heldout/xxx", "text": "..."}   (jsonl)  OR  {id: text} (json)
# 2. score (same normalization as the leaderboard):
pip install rapidfuzz
python score.py --hyps my_hyps.jsonl

Scoring / normalization

WER & CER over normalized text: combining marks + Quranic annotation signs + tatweel stripped; hamza forms and alef-maqsura unified. The alef-insensitive column additionally strips alef to neutralize the Uthmani-rasm vs imlaei spelling-convention mismatch (eval refs are rasm-ish; many models output imlaei), reflecting true word accuracy over orthography.

Inference config (language hint)

Each model is run to transcribe in Arabic, with the language set wherever the model exposes one, so every model is evaluated in its intended Arabic mode (not language-auto-detect, which under-rates multilingual models on short clips):

Model / family	Arabic hint
Whisper (large-v3)	`language="arabic"`, `task="transcribe"`
Seamless-m4t-v2	`tgt_lang="arb"`
omniASR	`lang="arb_Arab"`
Cohere transcribe	`language="ar"`
Gemma-4	prompt "transcribe in Arabic"
Qwen3-ASR	`language="Arabic"`
Tarteel (official)	`ar-SA`
ElevenLabs Scribe v2	`ar`
Deepgram nova-3	`ar`
Speechmatics enhanced	`ar`
Google Chirp 3 (sync + realtime)	`ar-XA`
Arabic-only models (this/nvidia-ar/Tarteel-whisper)	n/a (single-language)

Decoding is greedy / each model's default. No language-model rescoring, and no forced alignment - every model only ever sees the audio, never the reference transcript. Whisper uses short-form decoding (timestamp tokens only for the rare clip > 30 s, which Whisper requires for long audio; this is the model's own prediction, not alignment to the reference).

Leakage control

The public EveryAyah test split overlaps most Quran ASR training sets, so we use only the EveryAyah reciters with zero training overlap, plus a held-out QUL reciter and held-out real-phone tlog. nvidia FastConformer (current public #1) is scored on the same clips for a fair head-to-head.

Acknowledgements

Special thanks to David Jiménez Paredes / deepdml for the original community effort around the Open Universal Arabic Quranic ASR Leaderboard, which helped inspire this benchmark and leaderboard work. His feedback directly improved this dataset's Hugging Face structure (source-specific configurations and direct transcript fields for easier downstream evaluation), and his careful review materially raised its rigor - catching a scorer/results metric mismatch, identifying that whisper-large-v3 was being evaluated without the Arabic language hint, and contributing the prediction script used to standardize inference across models.

Provenance & licensing

Audio is redistributed for research reproducibility under a manual-approval gate. Upstream sources and their terms: EveryAyah (tarteel-ai/everyayah, CC-BY 4.0), tlog (tarteel-ai/tlog, gated), QUL (qul.tarteel.ai). Respect each upstream license; do not redistribute the audio. Reciters and tlog users retain their rights. If you are a rights-holder and want a clip removed, open a discussion.

Revision v1.1 (2026-08-15): corrected references for 10 qul_alnufais clips

Ten clips in the qul_alnufais split contain more speech than their reference covered. In nine of them the reciter recites the ayah twice in the same recording; in one (51_56) the recording contains 51:56, then 51:57, then 51:56 again.

The defect originates upstream, not in this benchmark's pipeline or in any evaluated model. The audio files were obtained from QUL (qul.tarteel.ai), already segmented per ayah and named surah_ayah by QUL; QUL presents these segments as human-reviewed. Each reference here is the canonical text of exactly the ayah QUL's own file naming declares. For these ten files the audio as distributed by QUL contains repeated recitation that the per-ayah naming does not reflect, so any faithful transcription of the full recording disagreed with the reference.

The affected clips: 37_43, 37_83, 24_20, 50_21, 43_46, 6_132, 6_102, 17_72, 6_32, 51_56.

Each correction was verified against the audio before being applied: the repeated composition was selected by matching a phoneme-level decode of the recording against candidate reference compositions (for 51_56 the chosen composition matches the audio at normalized distance 0.000 where the single-ayah reference does not). No audio was changed and no clips were removed; only reference_text for these 10 rows.

Impact on scores: models that transcribed the full recording faithfully were penalized under v1.0 and improve under v1.1; models that truncated after the first recitation lose the advantage the defect gave them. All prediction files with stored hypotheses (24 models) were rescored and the leaderboard updated. Entries whose hypotheses are no longer available are marked as scored against v1.0 references. For most models the two scales differ by about 1 WER point or less.