You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Quran Tajweed Benchmark

A test for systems that listen to Quran recitation. Two questions: does the system leave correct recitation alone, and does it catch the mistakes? Quran Lab runs every system on the leaderboard itself, on the same audio.

Leaderboard: Quran Tajweed Leaderboard

What is public, and what is not

  • Public, here: correct recitation. 3,910 recordings, 9.2 hours, each with the mushaf text, the phonemes and every legal way of reading its ayah. Most are by 22 professional reciters whose recordings were held out of training.
  • Private: the recitations with mistakes. The leaderboard also scores every system on recordings with one known mistake, in the studio and through a phone. Those recordings are not published, so that nobody can train on them. Send a model and Quran Lab runs it on them (see below).

Configs

All audio is 16 kHz mono FLAC. Every config has a single split, test.

config recordings hours what it is
correct_t1p_dev 930 2.20 Held-out panel, development half: 6 professional reciters. Use it to tune.
correct_t1p_test 1,780 4.24 Held-out panel, test half: 16 professional reciters. The leaderboard's correct-recitation numbers are measured on these recordings.
correct_t1 153 0.34 An older held-out reciter (Ali Hussein Qasim).
correct_t1x 1,044 2.37 4 reciters whose voices are in Quran Lab's own training mix, marked so: held out for other models, not for ours.
correct_t1w_test 3 0.03 Women: published competition recordings held out of every training pool (2 reciters, 3 recordings: all that passed every check).

Columns

  • audio: the recording, 16 kHz FLAC.
  • ayah, uthmani: the ayah (surah:ayah) and its Uthmani text.
  • ref_tj3: what the audio says, in phoneme tokens with | between words. canonical_tj3: the written text in the same tokens. alt_tj3: other accepted references.
  • lattice (JSON): every legal reading of the ayah: each word's forms for continuing (wasl) and stopping (waqf), starting after a stop with its pausal forms (sukun, rawm, ishmam), and both readings at disputed sites. A model is never charged for a legal stop or a legal reading.
  • labels (JSON): madd lists every madd with its legal lengths; sites is empty for correct recitation.
  • clip_is_error, identity_control, n_error_sites, error_classes: all false, zero or empty here; they keep the columns the same as in the private set.
  • speaker_id, speaker_name, gender, age_group, voice_status, channel, source_dataset, source_url.
  • seen_by_b7_models: the recording is a near-copy of audio in one of Quran Lab's older training mixes.
  • certification (JSON): how the row was certified, check by check.

How systems are scored

  • WER on correct recitation, studio and through a phone. Phoneme models are turned into text with quran-p2g, not told the ayah, so phoneme and text models are measured the same way.
  • PER for phoneme models, against the closest legal reading: a legal stop or reading is never an error.
  • Checker score, the leaderboard's ranking: of 100 recitations, half correct and half with one mistake, how many get exactly the right verdict (the correct ones left alone, the mistaken word pointed to). Never saying anything scores 50, marking everything 0. The half with mistakes is measured on the private recordings.

Use it

from datasets import load_dataset
correct = load_dataset("Quran-Lab/quran-tajweed-benchmark", "correct_t1p_test", split="test")

The scorer is in scorer/: parquet_to_bench.py writes a config out as files, and score_v4_bench.py scores phoneme or text output against it.

Get on the leaderboard

Put your model on the Hub (or send an ONNX export) with a script that reads 16 kHz audio, and open a discussion on the leaderboard Space. Quran Lab runs it on the private recordings and publishes the numbers.

Good to know

  • Recordings by women are few (three), and there are none by children yet.
  • EveryAyah recordings are not used: EveryAyah is in nearly every Quran model's training data, so a clip from it cannot show how a model handles audio it never heard.
  • These recordings are excluded from all of Quran Lab's future training. If you train on any of this data, say so when you report numbers.

License and sources

Quran-Lab No-Profit License, version 1.2 (NPL-1.2). See LICENSE. The audio comes from published recitations; the source_url column gives the original recording where there is one.

Citation

@misc{quranlab_tajweed_benchmark_2026,
  title  = {Quran Tajweed Benchmark},
  author = {Quran Lab},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Quran-Lab/quran-tajweed-benchmark}
}
Downloads last month
119

Spaces using Quran-Lab/quran-tajweed-benchmark 2