quranlab/quran-audio-text

QuranLab — Verse-Aligned Quran Text + Recitation References

This dataset joins QuranLab's canonical Hafs Arabic text to its per-ayah recitation references. Every row is one exact (recitation_id, verse_key) pair: the Uthmani transcript, a search-friendly Simple-Clean transcript, and the corresponding audio_url.

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 text and the recordings it is paired with reach you through the work of Tanzil Project, everyayah.com, 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.

What “combined” means

• The default manifest config has 268,148 rows: 43 Hafs recordings × 6,236 ayahs.
• Every per-recitation config has exactly 6,236 rows in canonical order, from 1:1 through 114:6.
• Audio fields are copied from quranlab/quran-audio without changing their URLs, reciter identity, bitrate, provenance, or word timing.
• text is verbatim Tanzil Uthmani v1.1; text_simple is verbatim Tanzil Simple-Clean v1.1.

Audio bytes are not hosted. audio_url points to the original public source, exactly as in the companion audio dataset — fetch the audio from there.

Why only Hafs

The text spine is Hafs an Asim with the Kufan 6,236-ayah enumeration. The three per-ayah Warsh recordings in the audio catalog use a different reading and enumeration basis, so pairing them with Hafs text would not be exact. They are intentionally excluded here and remain available in quranlab/quran-audio`. They can be added only after a verified, matching Warsh text spine is available.

Quick start

from datasets import load_dataset

# All verified Hafs text-audio pairs.
pairs = load_dataset(
    "quranlab/quran-audio-text", "manifest", split="train"
)
print(pairs[0]["verse_key"], pairs[0]["text"], pairs[0]["audio_url"])

# One reciter: exactly 6,236 rows, with the full audio metadata/timing schema.
husary = load_dataset(
    "quranlab/quran-audio-text", "husary", split="train"
)

To retrieve audio, fetch audio_url from its original host.

Terms and credit for the text and the recordings are identical for every row, so they are published once per config in metadata/configs.csv instead of being repeated on all 268,148 rows.

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 plus the cross-recording manifest
• Join key: (recitation_id, verse_key) is unique per row; verse_key joins to quranlab/quran, recitation_id to quranlab/quran-audio
• 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 source.
• Filter on timing_quality == "strict" for word-level training. Rows labelled source-artifact carry inherited alignment imperfections and are marked rather than silently mixed in.
• Every row is Hafs; divergent readings are excluded here on purpose, so text and audio cannot disagree on verse division.

Inspect it without downloading the data:

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

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

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

Query it in place with SQL — no full download:

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

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

Important columns

Column	Meaning
verse_key	Canonical key, such as 2:255
text	Verbatim Uthmani Arabic transcript
text_simple	Verbatim Simple-Clean Arabic transcript
audio_url	Matching per-ayah recording reference
recitation_id	Stable recording/config identifier
riwayah	Always hafs-asim in this strict dataset
has_word_timing	Whether this verse has inherited word segments
segments	Word timing in per-recitation configs when available
timing_quality	strict, source-artifact, or missing; filterable alignment safety label

Configs

The default manifest is convenient for ASR/TTS indexing and streaming. Individual recording configs preserve the complete schema of the corresponding audio config:

config	reciter	style	strict timing	source artifacts	missing timing
abdul-basit	Abdul Basit Abdus Samad	murattal	6,200	36	0
abdul-basit-mujawwad	Abdul Basit Abdus Samad	mujawwad	6,106	130	0
abdul-rahman-al-sudais	Abdul Rahman Al-Sudais	murattal	6,236	0	0
ali-al-hudhaify	Ali Al-Hudhaify	murattal	6,236	0	0
husary	Mahmoud Khalil Al-Husary	murattal	6,215	21	0
husary-muallim	Mahmoud Khalil Al-Husary	muallim	6,206	30	0
husary-mujawwad	Mahmoud Khalil Al-Husary	mujawwad	6,236	0	0
maher-al-muaiqly	Maher Al-Muaiqly	murattal	6,236	0	0
mahmoud-ali-al-banna	Mahmoud Ali Al-Banna	murattal	6,235	0	1
minshawi	Mohamed Siddiq El-Minshawi	murattal	6,199	37	0
minshawi-mujawwad	Mohamed Siddiq El-Minshawi	mujawwad	6,146	89	1
mishary-alafasy	Mishary Rashid Alafasy	murattal	6,206	30	0
muhammad-ayyub	Muhammad Ayyub	murattal	6,236	0	0
saad-al-ghamdi	Saad Al-Ghamdi	murattal	6,236	0	0
saud-al-shuraim	Saud Al-Shuraim	murattal	6,182	53	1
abdullah-al-juhani	Abdullah Awad Al-Juhany	murattal	6,236	0	0
abdullah-basfar	Abdullah Basfar	murattal	6,236	0	0
abu-bakr-al-shatri	Abu Bakr Al-Shatri	murattal	6,201	35	0
ahmad-al-ajmi	Ahmad Al-Ajmi	murattal	6,234	0	2
ali-jaber	Ali Abdullah Jaber	murattal	6,236	0	0
ayman-suwaid	Ayman Rushdi Suwaid	murattal	6,236	0	0
fares-abbad	Fares Abbad	murattal	6,236	0	0
hani-ar-rifai	Hani Ar-Rifai	murattal	6,194	42	0
khalifa-al-tunaiji	Khalifa Al-Tunaiji	murattal	6,236	0	0
mohamed-al-tablawi	Mohammad Al-Tablawi	murattal	6,180	56	0
muhammad-jibreel	Muhammad Jibreel	murattal	6,236	0	0
nasser-al-qatami	Nasser Al-Qatami	murattal	6,236	0	0
yasser-al-dosari	Yasser Al-Dosari	murattal	6,236	0	0
abdullah-matroud	Abdullah Al-Matrood	murattal	6,236	0	0
ahmed-neana	Ahmed Neana	murattal	6,235	0	1
akram-al-alaqimy	Akram Al-Alaqimy	murattal	6,236	0	0
ali-hajjaj-al-suesy	Ali Hajjaj Al-Suesy	murattal	6,236	0	0
aziz-alili	Aziz Alili	murattal	6,236	0	0
ibrahim-al-akhdar	Ibrahim Al-Akhdar	murattal	6,236	0	0
karim-mansoori	Karim Mansoori	murattal	6,236	0	0
khalid-al-qahtani	Khalid Abdullah Al-Qahtani	murattal	6,236	0	0
muhammad-abdulkareem	Muhammad Abdul Kareem	murattal	6,233	0	3
muhsin-al-qasim	Abdul Muhsin Al-Qasim	murattal	6,236	0	0
parhizgar	Shahriar Parhizgar	murattal	6,235	0	1
sahl-yassin	Sahl Yasin	murattal	6,236	0	0
salah-al-budair	Salah Al-Budair	murattal	6,236	0	0
salah-bukhatir	Salah Abdul Rahman Bukhatir	murattal	6,236	0	0
yaser-salamah	Yaser Salamah	murattal	6,236	0	0

Integrity and reproducibility

QUALITY_REPORT.md records the last full validation. The gate checks:

• exact 43 × 6,236 pair coverage and composite-key uniqueness;
• byte-for-byte text equality to both canonical text configs;
• column-for-column equality of inherited audio data;
• exact SSSAAA.mp3 URL construction for every pair;
• strict Hafs/Kufan enumeration and exclusion of divergent readings;
• complete per-config terms and credit in metadata/configs.csv;
• recomputed timing bounds/token coverage and the derived timing_quality label;
• upstream checksum-manifest coverage when validation is run with --verify-upstream.

Machine-readable source/output hashes and scope decisions are in metadata/build_manifest.json.

Word-timing quality

The original timing payload is preserved verbatim. QuranLab-generated timing passes the strict temporal/token-coverage gate. A small documented fraction of the inherited cpfair/quran-align rows contains skipped-token or non-monotonic alignment artifacts; these rows are explicitly labeled timing_quality=source-artifact instead of being silently presented as strict. Use timing_quality == "strict" for word-level training. This label does not affect the verse text or audio_url pairing.

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
Text is byte-identical to the canonical corpus	both transcripts are compared cell-by-cell against quranlab/quran on every build; a single differing character fails the release
Every pair is exact	268,148 rows = 43 complete recordings x 6,236 ayahs, with composite-key uniqueness asserted
Audio metadata is inherited, not re-derived	URLs, reciter identity, bitrate and timing are copied column-for-column from the audio corpus and checked for drift
Alignment quality is labelled honestly	timing_quality separates strict timing from inherited source artifacts instead of presenting everything as clean
Every referenced file exists upstream	268,148 filenames verified against upstream inventories without downloading audio

from datasets import load_dataset

ds = load_dataset("quranlab/quran-audio-text", "husary", split="train")
assert len(ds) == 6236

# the transcript really is the canonical text
q = load_dataset("quranlab/quran", "arabic-uthmani", split="train")
assert ds["text"] == q["text"]

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

Uthmani and Simple-Clean text: Tanzil Project, CC BY 3.0, kept verbatim. Word timing: CC BY 4.0. The recordings themselves are referenced by link, never hosted. Details in LICENSES.md.

Citation

@misc{quranlab_quran_audio_text_2026,
  title        = {QuranLab — Verse-Aligned Quran Text and Recitation References},
  author       = {QuranLab},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/quranlab/quran-audio-text}}
}