Quran-Lab/quran-tajweed-phonetics

The complete phonetic layer of the Quran in the riwaya of Hafs 'an 'Asim via tariq al-Shatibiyyah: 6,236 ayat, 522,475 phones, every phone carrying its tajweed attribution: madd class with its transmitted length range, ghunna grade, qalqalah class, tafkheem with its rank, sakt, the seventeen sifat, and the rule that produced it.

Built and maintained by Quran Lab, a waqf building open technology in the service of the Quran.

How it was built and verified

• Indexed from the books. Machine-assisted passes read the classical literature page by page, about 3,300 pages, and indexed every candidate ruling. The index is a card catalogue, not an authority: nothing enters the engine on a model's word.
• Page-verified citations. Every ruling's citation was checked against the prints themselves: the book, the volume:page or bayt, the subject at that location, and quoted matn lines verbatim. The full audit is committed in the engine repository.
• A gate-enforced register. The engine source, the rulings register, and the review sheet must agree with each other or the test suite fails. Five independent oracles, corpus-wide behavioral audits, and seeded-bug drills prove the engine implements the register faithfully.
• Blinded sanad review, complete. An ijazah-holding hafiz ruled on all 208 register entries from his own talaqqi: 191 confirmed, 17 confirmed while documenting a second transmitted wajh, zero rulings judged wrong. His independence was measured, not assumed: in the blind screening, planted errors dressed in the real citation format were all caught. Final tally below.
• Sealed at the top of the discipline. The complete register then underwent formal scholarly arbitration (تحكيم علمي) by a serving professor at the Department of Qira'at, Umm al-Qura University in Makkah, the city where the revelation began. He ruled on all 208 entries from what he received before his own teachers and shaykhs («وفق ما تلقيته وقرأته على أساتذتي ومشايخي», his words), blind to the first reviewer and his verdicts, and confirmed the register. At his own request he is credited by rank alone: the work needed his judgment, not his name. Two unbroken chains have now examined every ruling independently. Both reached the same answer.
• The method is the tradition's own. الرواية والدراية: transmission and verification, given machinery. Ibn al-Jazari built al-Nashr exactly this way: every written source gathered, every reading verified through living chains, nothing accepted without both. This project implements that method.

Backed by the books

This is not a heuristic transcription. Every rule in the generating engine is implemented directly from the classical literature and carries its citation with volume and page: the Shatibiyyah and its commentaries (فتح الوصيد، إبراز المعاني، سراج القارئ، الوافي), Ibn al-Jazari's al-Nashr, al-Tayyibah, and al-Muqaddimah, al-Dani's al-Taysir and Jami' al-Bayan, Tuhfat al-Atfal, and the tajweed masters through Hidayat al-Qari, around seventy works in total. Every khilaf the tradition transmits (يبصط، ضعف، سلاسلا، آتاني، the waqf raa'at and more) is carried as an explicit, cited choice, never a silent default. The full rulings register and bibliography live in the engine repository.

Validation, over the whole corpus:

• two independently pinned mushaf editions (Tanzil, KFGQPC) produce identical phones for all 6,236 ayat
• 100% agreement with the mushaf's own dabt at ~8,900 tanween sites
• checked against the quran-tajweed rule-span dataset (Dar al-Maarifah-derived, github.com/cpfair/quran-tajweed; 60,057 annotations): all phones correct. 60,056 annotations match; the single exception at 17:7 is a wrong annotation in that span file itself (it reads the bare rasm as a 2-count madd, while the rasm literature rules the recitation restores the elided waw and the madd is muttasil), so the phones here are right and the checker is wrong on that one entry
• every wasl-initial word cross-checked against the Quranic Arabic Corpus morphology, zero disagreements
• 552 tests, a 25-mutant harness drill, and a frozen determinism hash

The human expert review is complete (Rabi' al-Awwal 1448 / August 2026): each of the 208 entries in the engine's rulings register was independently ruled on (191 confirmed, 17 marked as carrying a second transmitted wajh, zero rulings judged wrong) by a board-examined hafiz (Wifaq al-Madaris al-Salafiyyah, mumtaz, 1443/2022) holding a written ijazah in Hafs 'an 'Asim via tariq al-Shatibiyyah (1434/2013), its text carrying the authorization to transmit and a connected sanad to the Prophet ﷺ, with a full-Quran tajweed certificate (mumtaz, 377/400) on the classical curriculum, certified by the Prophet's Mosque Qur'an program and currently a Shari'ah student at the Islamic University of Madinah. Credential scans are on file with the maintainer; the reviewer is credited, at his own preference, as Shaikh Sami Almadani. Before engagement he passed a blind screening of ten rulings containing three deliberately planted errors: all three caught, zero false alarms. The same screening has since failed other candidate reviewers, which is what makes passing it meaningful. Every disagreement was adjudicated against the cited classical texts on the public record (ADJUDICATIONS.md): the review's catches were upheld, credited to the reviewer verbatim, proved to be labeling slips rather than engine errors, and the defect class they exposed was audited out of all 208 rows and closed with a permanent test gate. Above this review, formal scholarly arbitration (تحكيم علمي) of the complete register has been completed by a serving professor at the Department of Qira'at, Umm al-Qura University in Makkah, ruling on every entry from his own learning with no contact with the first reviewer; his objections were adjudicated on the same public record (entries 5 and 6), and at his own request he is credited by academic title only. Two independent chains of transmission, with no contact between them, have now each confirmed the register in full. Final tally, generated live from the review sheet:

Acknowledgement

The pioneering open tajweed phonetizer is quran-transcript (on Hugging Face). Our engine was built clean-room from the classical sources and shares none of its code, but quran-transcript served as the differential baseline of our validation, agreeing with it at 98.9% character-level across the whole corpus. May Allah reward its author.

Files

file	contents
quran_phonetics.jsonl	the flagship: one row per ayah, every phone fully attributed
quran_labels_v1.jsonl	training labels: length-tagged tajweed tokens plus a legacy repeat-string encoding
ayah_tokens.jsonl	the token sequences alone
tokens.txt, vocab_manifest.json	the 234-token tajweed vocabulary (hash-manifested, blank last)
rule_index.jsonl	per rule-bearing phone: rule id, length sets, metadata
rulings.jsonl	the 62 engine rulings: Arabic and English name, phase, and the classical citation by book and volume:page. Every rule_id that appears on a phone resolves here to its source in the books
register_reviewed.jsonl	the complete reviewed register: all 208 rulings with their ayah witness, wording, citation, and the expert review verdict on each row (191 صحيح, 17 فيه وجهان documenting a second transmitted wajh), reviewed line by line by Shaikh Sami Almadani (hafiz; written ijazah with connected sanad) and confirmed in formal arbitration by a serving professor of Qira'at at Umm al-Qura University in Makkah
bijection_old250.json	warm-start map onto a legacy 250-unit chunk vocabulary

Phone schema

{
  "surah": 1, "ayah": 1, "text": "...",
  "phones": [
    {
      "base": "sad", "kind": "consonant", "geminated": false,
      "word_index": 3, "tafkheem": "mofakham", "tafkheem_rank": 2,
      "qalqalah": "sughra",
      "length": {"kind": "free", "allowed": [2, 4, 6], "canonical": 4,
                 "scoring": [2, 4, 6]},
      "rules": ["R144_IKHFA"],
      "sifat": {"hams_or_jahr": "jahr", "...": "..."}
    }
  ]
}

Conventions that matter:

• Lengths are set-valued prescriptions, not observations. allowed is the Shatibiyyah-legal range, scoring the attested superset, canonical a deterministic default. Where the tradition gives the reciter a choice (munfasil, 'aared, leen), the value actually realized in a recording is a measurement this text layer cannot supply. Do not treat canonical values as ground truth at free-choice positions.
• Rows are phonemized at ayah-end waqf. Mid-ayah pause variants, rawm and ishmam alternates, and junction-true multi-ayah wasl labels can all be generated with the engine.

License

Quran-Lab No-Profit License 1.1 (see LICENSE): free for any non-profit use, share-alike, and no profit may be obtained from it or from derivatives, including models trained with it.

This dataset states the prescription of the riwaya; certifying reciters remains the sanad's work.

Citation

@software{quran_g2p,
  author = {{Quran Lab}},
  title  = {quran-g2p: a specification-first tajweed phonemizer for
            Hafs 'an 'Asim (tariq al-Shatibiyyah)},
  year   = {2026},
  url    = {https://github.com/Muno459/quran-g2p}
}

وَرَتِّلِ الْقُرْآنَ تَرْتِيلًا