quranlab/quran-audio

Dataset Card for QuranLab — Qur'an Recitation Audio

A verse-aligned reference layer for Qur'an recitation audio: a unified taxonomy of 245 reciters, per-ayah and per-surah reference manifests that link every verse to its original public source, and word-precise timing released under CC-BY-4.0. This dataset hosts no audio files — it points to each recitation where it already lives and contributes the connective scholarship: taxonomy, alignment, and timing.

QuranLab is a volunteer effort. Our aim is to present these works carefully and at high quality, and to help them travel faithfully — in the spirit in which they were written — not to claim them as ours.

The recordings referenced here, and the timing published with them, reach you through the work of everyayah.com, mp3quran.net, cpfair/quran-align and jonatasgrosman/wav2vec2-large-xlsr-53-arabic. Who to thank, and what we consulted without ever quoting, is set out in SOURCES.md.

It is the audio companion to the text corpus quranlab/quran (Arabic text, translations, tafsīr, word morphology); the two join on the column verse_key.

---

What's here — and what's not

This dataset does not host any audio files, and that is intentional.

A recitation is the reciter's performance and the producer's master, so we don't re-host it. For each unit of recitation we publish a reference audio_url pointing to the original public source — fetch it from there — alongside our own word timing.

Included (openly)	Referenced (not hosted)
Unified 245-reciter taxonomy	The recitation audio bytes (© reciters & producers)
Per-ayah & per-surah manifests with an external audio_url	— fetch them from the original source via the URL we provide
CC-BY-4.0 word timing for 44 recordings	—

---

Quick start

from datasets import load_dataset

# Default: the per-ayah reference index (every recording × ayah, with audio_url)
manifest = load_dataset("quranlab/quran-audio")

# The unified 245-reciter taxonomy
reciters = load_dataset("quranlab/quran-audio", "reciters",       split="train")

# The per-surah breadth layer (285 recordings across 15 riwāyāt, mp3quran)
surahs   = load_dataset("quranlab/quran-audio", "surah-manifest", split="train")

# A single recording, with its CC-BY word-level timing
husary   = load_dataset("quranlab/quran-audio", "husary",         split="train")

To pull text alongside audio, load the companion corpus quranlab/quran and join on verse_key. To listen to or process the audio, fetch each file yourself from the audio_url it points to, and honor the rights of the reciter and producer.

---

What we contribute

This dataset is more than an aggregation. Its heart is three pieces of original, freely-given work.

1. A unified reciter taxonomy — 245 reciters. Names in Arabic and English, nationality, era, a popularity tier, and the riwāyāt (recitation styles) each reciter offers. To our knowledge, no prior public audio set offered this clean, cross-source map of who recites what, in which reading.
2. 33 recordings of brand-new word timing. Open word-level timing previously existed for only 11 recordings, so we generated our own for 33 more and released it under CC-BY-4.0 — bringing freely-usable word timing to 44 recordings in total.
3. Truthful coverage marking. Where timing or a source is missing, the row says so rather than guessing.

---

Dataset Details

The dataset is organized as a small set of cleanly joinable layers:

• A per-ayah reference manifest — 47 recordings (from everyayah.com), each keyed to the canonical 6,236-ayah numbering (Ḥafṣ ʿan ʿĀṣim) via verse_key, with an external audio_url. This is the precise, ayah-by-ayah layer and the default configuration.
• A per-surah reference layer — 285 recordings (from mp3quran.net) spanning 15 qirāʾāt / riwāyāt — the breadth layer: hundreds of reciters and all the major readings.
• Word-precise timing (CC-BY-4.0) for 44 of the 47 per-ayah recordings — of these, 11 come from the community project cpfair/quran-align and 33 we generated ourselves.

Dataset Sources

• Repository: https://huggingface.co/datasets/quranlab/quran-audio
• Companion text corpus: https://huggingface.co/datasets/quranlab/quran (join on verse_key)
• Per-ayah audio references: everyayah.com (VerseByVerseQuran.com)
• Per-surah audio references: mp3quran.net
• Upstream open timing (11 recordings): cpfair/quran-align (© Collin Fair, CC-BY-4.0)
• Reference text for alignment: the Tanzil project (Uthmani text)

---

For agents and pipelines

Deterministic facts an automated consumer needs, so nothing has to be inferred from a sample.

• Config naming: one config per recording (husary, abdul-basit, …), plus manifest, surah-manifest and reciters
• Join key: verse_key to the text corpus; reciter_id to the reciters taxonomy; segments[].word_position to word-level morphology
• Default config: manifest
• Format: Parquet, one directory per config, split train.
• Terms and credit are per config in metadata/, never per row — join on the config name when you need them.
• No audio bytes are hosted. audio_url points at the original public source; fetch it from there under that host's terms.
• Gate on has_word_timing per row — coverage is per ayah, not per recording.
• Keep readings separate: riwayah is on every row and Warsh recordings do not share the Hafs verse division.

Inspect it without downloading the data:

# every config and split
curl https://datasets-server.huggingface.co/splits?dataset=quranlab/quran-audio

# a first page of rows
curl "https://datasets-server.huggingface.co/first-rows?dataset=quranlab/quran-audio&config=husary&split=train"

# machine-readable schema and provenance (Croissant JSON-LD)
curl https://huggingface.co/api/datasets/quranlab/quran-audio/croissant

Query it in place with SQL — no full download:

hf datasets sql quranlab/quran-audio "SELECT * FROM 'husary' LIMIT 5"

import duckdb
duckdb.sql("SELECT count(*) FROM 'hf://datasets/quranlab/quran-audio/husary/*.parquet'")

Dataset Structure

The dataset ships several configurations. The per-ayah manifest is the default.

Config	What it is
manifest (default)	Per-ayah reference index: every recording × ayah, keyed to the canonical 6,236-ayah numbering via verse_key, each with an external audio_url. Covers all 47 recordings (everyayah.com).
reciters	The unified 245-reciter taxonomy (one row per reciter).
surah-manifest	Per-surah breadth layer: 285 recordings (mp3quran.net) across 15 qirāʾāt / riwāyāt, organized per surah, each with an external audio_url.
<recitation_id>	One config per per-ayah recording (47 in all, e.g. husary, mishary-alafasy). Each has 6,236 rows; the 44 timed recordings additionally carry a segments word-timing column.

Fields (per-recitation / per-ayah configs)

Field	Type	Description
verse_key	string	Canonical surah:ayah key (Ḥafṣ ʿan ʿĀṣim, 6,236 ayahs); the join key to quranlab/quran.
surah, ayah	int	Surah and ayah numbers.
recitation_id, reciter_id	string	Recording and reciter identifiers (the latter joins to the reciters taxonomy).
name_en	string	Reciter name (cross-recitation manifest and surah-manifest only, where it varies row to row).
riwayah, style	string	Reading (e.g. hafs-asim) and style (murattal / mujawwad / muallim).
audio_url	string	Reference link to the original audio at its public source (no bytes stored here).
has_word_timing	bool	Whether word timing is present for this row.
segments	list	Per-word timing: each entry has word_position (1-based; joins to the word-morphology layer of quranlab/quran), start_ms, and end_ms within the ayah audio.
timing_source	string	Which alignment produced the timing for this row (cpfair/quran-align or ours).

Terms, credit and source links are the same for every ayah of a recording, so they are published once per config in metadata/configs.csv rather than repeated on all 6,236 rows. Reciter details live in the reciters config.

Illustrative timed row (husary, ayah 1:2):

{
  "verse_key": "1:2",
  "surah": 1,
  "ayah": 2,
  "recitation_id": "husary",
  "reciter_id": "husary",
  "riwayah": "hafs-asim",
  "style": "murattal",
  "audio_url": "https://everyayah.com/data/Husary_128kbps/001002.mp3",
  "has_word_timing": true,
  "segments": [
    {"word_position": 1, "start_ms": 700,  "end_ms": 1410},
    {"word_position": 2, "start_ms": 1420, "end_ms": 2200}
  ],
  "timing_source": "cpfair/quran-align"
}

The reciters config carries, per reciter: name (Arabic and English), nationality, era, a popularity tier, riwayat_available, and styles_available. The surah-manifest config mirrors the per-ayah fields at surah granularity, with a surah_audio_url. The taxonomy unifies naming across everyayah.com and mp3quran.net, so a single reciter is no longer duplicated under inconsistent spellings — making it straightforward to select reciters by reading, region, or era.

Per-ayah recordings (47) — reciter, riwāyah, style, and word-timing source

recording (config)	reciter	riwāyah	style	word timing
abdul-basit	Abdul Basit Abdus Samad	hafs-asim	murattal	cpfair (CC-BY)
abdul-basit-mujawwad	Abdul Basit Abdus Samad	hafs-asim	mujawwad	cpfair (CC-BY)
abdul-basit-warsh	Abdul Basit Abdus Samad	warsh-nafi	murattal	— (reference-only)
abdul-rahman-al-sudais	Abdul Rahman Al-Sudais	hafs-asim	murattal	QuranLab (CC-BY)
ali-al-hudhaify	Ali Al-Hudhaify	hafs-asim	murattal	QuranLab (CC-BY)
husary	Mahmoud Khalil Al-Husary	hafs-asim	murattal	cpfair (CC-BY)
husary-muallim	Mahmoud Khalil Al-Husary	hafs-asim	muallim	cpfair (CC-BY)
husary-mujawwad	Mahmoud Khalil Al-Husary	hafs-asim	mujawwad	QuranLab (CC-BY)
maher-al-muaiqly	Maher Al-Muaiqly	hafs-asim	murattal	QuranLab (CC-BY)
mahmoud-ali-al-banna	Mahmoud Ali Al-Banna	hafs-asim	murattal	QuranLab (CC-BY)
minshawi	Mohamed Siddiq El-Minshawi	hafs-asim	murattal	cpfair (CC-BY)
minshawi-mujawwad	Mohamed Siddiq El-Minshawi	hafs-asim	mujawwad	cpfair (CC-BY)
mishary-alafasy	Mishary Rashid Alafasy	hafs-asim	murattal	cpfair (CC-BY)
muhammad-ayyub	Muhammad Ayyub	hafs-asim	murattal	QuranLab (CC-BY)
mustafa-ismail	Mustafa Ismail	hafs-asim	murattal	QuranLab (CC-BY)
saad-al-ghamdi	Saad Al-Ghamdi	hafs-asim	murattal	QuranLab (CC-BY)
saud-al-shuraim	Saud Al-Shuraim	hafs-asim	murattal	cpfair (CC-BY)
abdullah-al-juhani	Abdullah Awad Al-Juhany	hafs-asim	murattal	QuranLab (CC-BY)
abdullah-basfar	Abdullah Basfar	hafs-asim	murattal	QuranLab (CC-BY)
abu-bakr-al-shatri	Abu Bakr Al-Shatri	hafs-asim	murattal	cpfair (CC-BY)
ahmad-al-ajmi	Ahmad Al-Ajmi	hafs-asim	murattal	QuranLab (CC-BY)
ali-jaber	Ali Abdullah Jaber	hafs-asim	murattal	QuranLab (CC-BY)
ayman-suwaid	Ayman Rushdi Suwaid	hafs-asim	murattal	QuranLab (CC-BY)
fares-abbad	Fares Abbad	hafs-asim	murattal	QuranLab (CC-BY)
hani-ar-rifai	Hani Ar-Rifai	hafs-asim	murattal	cpfair (CC-BY)
khalifa-al-tunaiji	Khalifa Al-Tunaiji	hafs-asim	murattal	QuranLab (CC-BY)
mohamed-al-tablawi	Mohammad Al-Tablawi	hafs-asim	murattal	cpfair (CC-BY)
muhammad-jibreel	Muhammad Jibreel	hafs-asim	murattal	QuranLab (CC-BY)
nasser-al-qatami	Nasser Al-Qatami	hafs-asim	murattal	QuranLab (CC-BY)
yasser-al-dosari	Yasser Al-Dosari	hafs-asim	murattal	QuranLab (CC-BY)
abdullah-matroud	Abdullah Al-Matrood	hafs-asim	murattal	QuranLab (CC-BY)
ahmed-neana	Ahmed Neana	hafs-asim	murattal	QuranLab (CC-BY)
akram-al-alaqimy	Akram Al-Alaqimy	hafs-asim	murattal	QuranLab (CC-BY)
ali-hajjaj-al-suesy	Ali Hajjaj Al-Suesy	hafs-asim	murattal	QuranLab (CC-BY)
aziz-alili	Aziz Alili	hafs-asim	murattal	QuranLab (CC-BY)
ibrahim-al-akhdar	Ibrahim Al-Akhdar	hafs-asim	murattal	QuranLab (CC-BY)
ibrahim-al-dosary-warsh	Ibrahim Al-Dosary	warsh-nafi	murattal	— (reference-only)
karim-mansoori	Karim Mansoori	hafs-asim	murattal	QuranLab (CC-BY)
khalid-al-qahtani	Khalid Abdullah Al-Qahtani	hafs-asim	murattal	QuranLab (CC-BY)
muhammad-abdulkareem	Muhammad Abdul Kareem	hafs-asim	murattal	QuranLab (CC-BY)
muhsin-al-qasim	Abdul Muhsin Al-Qasim	hafs-asim	murattal	QuranLab (CC-BY)
parhizgar	Shahriar Parhizgar	hafs-asim	murattal	QuranLab (CC-BY)
sahl-yassin	Sahl Yasin	hafs-asim	murattal	QuranLab (CC-BY)
salah-al-budair	Salah Al-Budair	hafs-asim	murattal	QuranLab (CC-BY)
salah-bukhatir	Salah Abdul Rahman Bukhatir	hafs-asim	murattal	QuranLab (CC-BY)
yaser-salamah	Yaser Salamah	hafs-asim	murattal	QuranLab (CC-BY)
yassin-al-jazaery-warsh	Yassin Al-Jazaery	warsh-nafi	murattal	— (reference-only)

---

Uses

Direct Use

• Arabic speech recognition (ASR) — verse-aligned references and word timing for training and evaluation against verbatim text.
• Text-to-speech (TTS) and prosody / pronunciation modeling, using the recited Qur'an as a richly structured speech domain.
• Audio–text word-alignment research — word-level timings across many reciters and recording conditions.
• Qur'an applications and learning tools — highlighting each word as it is recited, tajwīd instruction, and memorization aids.

Because the companion corpus quranlab/quran joins on verse_key, timing can be mapped straight onto translations and per-word morphology.

Out-of-Scope Use

• Redistributing the recitation audio — get it from the original source instead; we only point at it.
• Any use that misrepresents the Qur'anic text or a reciter's performance.

---

Dataset Creation

Curation Rationale

Open work on the recited Qur'an has long been held back by two gaps: the absence of a clean taxonomy spanning the independent public audio sources, and the scarcity of open word timing. We set out to close both — to organize what already exists, and to contribute what did not.

Source Data

Data Collection and Processing

Reference manifests were assembled from the public catalogs of everyayah.com (per-ayah) and mp3quran.net (per-surah), with reciter identities reconciled into the unified taxonomy. For each recording we recorded the canonical metadata needed to align it to the text and attribute it correctly, and stored an external audio_url back to the original source. No audio bytes were retained. Where audio was needed to compute timing, each file was processed in memory and immediately discarded.

Who are the source data producers?

The recitations are the work of the reciters who performed them and the producers who recorded them; the recordings remain their property. The Uthmani text follows the Tanzil project. The original open timing for 11 recordings is the work of cpfair/quran-align.

Annotations — Word Timing

For the 33 recordings we contribute, we used forced alignment — lining up each ayah's verbatim Uthmani text with that ayah's recitation audio to mark where each word begins and ends — using only permissively-licensed open tools, and released the result under CC-BY-4.0. The output was validated for quality: timings are smooth, in word order, and free of gross artifacts. Coverage is recorded truthfully per row via has_word_timing.

Personal and Sensitive Information

The dataset contains the Qur'anic text, public metadata about reciters and recordings (names, nationality, era), external reference URLs, and word-timing offsets. It contains no private personal data.

---

Bias, Risks, and Limitations

• Reading coverage. The per-ayah timing layer is overwhelmingly Ḥafṣ ʿan ʿĀṣim (44 of 47 recordings). The 3 Warsh recordings follow a slightly different verse division, are kept as separate configs, and are reference-only (no timing). Treat Warsh as a distinct reading and take care with cross-reading joins.
• Per-ayah timing gaps. A few individual ayahs are untimed where the original audio file was missing or corrupt upstream; always consult has_word_timing rather than assuming full coverage.
• Forced-alignment imperfection. Automatically produced timing, however carefully validated, can contain occasional imperfect spans. It is well suited to research and tooling, not as infallible ground truth.
• External dependencies. Because audio is referenced rather than hosted, availability depends on the original sources; URLs may change or go offline over time.

Recommendations

Use has_word_timing to gate timing-dependent workflows, and keep Ḥafṣ and Warsh material separate.

---

Quality, and how to check it yourself

Nothing is released until the integrity gate passes. Every claim below is something you can re-run rather than take on trust — here on the published data, not in our build tree:

What is guaranteed	Evidence
Word timing we generated is openly licensed	33 recordings force-aligned by us and released CC-BY-4.0, on top of the 11 that already had open timing
Coverage is stated per row, not rounded up	has_word_timing is set from the actual segments present for that ayah, so a partially aligned recording is never presented as complete
Every reference resolves	all 268,148 per-ayah URL filenames were checked against the upstream inventories without downloading audio
One reciter, one identity	245 reciters reconciled across independent public catalogues, so the same reciter is not duplicated under variant spellings
Alignment is verifiable, not asserted	timings are monotonic, in word order and bounded by the ayah; known imperfect spans are labelled rather than smoothed over

from datasets import load_dataset

m = load_dataset("quranlab/quran-audio", "manifest", split="train")
assert "audio" not in m.column_names           # no bytes hosted, by design

h = load_dataset("quranlab/quran-audio", "husary", split="train")
assert len(h) == 6236
# timing is monotonic within each ayah
seg = next(r["segments"] for r in h if r["has_word_timing"])
assert all(a["end_ms"] <= b["start_ms"] for a, b in zip(seg, seg[1:]))

Where a source is genuinely missing, the row is published empty and labelled rather than filled by guessing from a neighbouring entry. Coverage numbers in this card count only real content.

Terms

The word timing is CC-BY-4.0 — free to use with credit to cpfair/quran-align for its 11 recordings and to QuranLab for the 33 we generated. The audio itself is referenced, not hosted: audio_url points at the original source, so fetch it from there. Details in LICENSES.md.

Reciters, producers, and their representatives: if you would like a reference row removed, open a request on the Community tab and we will handle it promptly.

---

The QuranLab family

These datasets are built by the same volunteers and are meant to be used together; they join on shared keys — verse_key on the Qur'an side, hadith_key on the Sunnah side.

Dataset	What it holds
quranlab/quran	Qur'an text, multilingual translations, tafsir, word-level morphology
quranlab/quran-audio	Recitation reference manifest and CC-BY word timing (references only, no audio files)
quranlab/hadith	Nine Books Arabic matn, translations, and normalized authenticity grades
quranlab/islamic-corpus-graph	The above linked into one structured graph, with retrieval passages and a held-out evaluation set

You are reading the card for quranlab/quran-audio. Corrections, additions, and removal requests are welcome on any of them through the Community tab.

---

Acknowledgments

This work stands on the generosity of others, and we thank them sincerely:

• everyayah.com (VerseByVerseQuran.com) and mp3quran.net, whose public services make the recited Qur'an widely available and which we reference here with gratitude.
• cpfair/quran-align (© Collin Fair) for openly-licensed (CC-BY-4.0) word timing covering 11 recordings — work that inspired our own contribution.
• The Tanzil project, for the Uthmani text used as the alignment reference.
• The maintainers of the open-source acoustic-model and alignment tooling that made our own word timing possible, and the communities behind them.

Above all, we acknowledge the reciters and producers whose recitations are the reason this project exists. The recordings remain entirely theirs; our work is only the connective layer around them.

---

Citation

@misc{quranlab_quran_audio_2026,
  title        = {QuranLab — Qur'an Recitation Audio: Reference Manifest and CC-BY Word Timing},
  author       = {QuranLab},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/quranlab/quran-audio}}
}

Dataset Card Contact

Questions, corrections, contributions, and rights-holder requests are warmly welcomed on the dataset's Community (Discussions) tab on the Hugging Face Hub.

Offered freely as a service to the worldwide community of learners, developers, and reciters — in the hope that it makes the words of the Qur'an a little easier to study, build with, and share. May it be of benefit.