nurlingo/quran-recitation-detect-split

Quran Recitation Detect & Split

258 human-reviewed ASR transcripts of real Quran recitations (254 Quranic, 4 non-Quran), with gold labels for two tasks:

1. Detection — which ayahs (verses) does the transcript contain?
2. Split — segment the transcript text by ayah.

Transcripts come from a production recitation-checking application: users recite from memory, audio is transcribed by a commercial ASR model, and each transcript was manually reviewed (ayah assignment, per-ayah split, confidence). Transcripts are noisy — they contain ASR errors, hesitations, repeated words, and openings such as the isti'adhah and basmala that are not part of the labeled ayahs. The 4 non-Quran rows are negative examples on which a system should abstain.

This is the frozen dataset used in the paper Autoresearch with Coding Agents: Generalizers and Metric-Maximizers on Quran Recitation Data (under review). The train/test split is exactly the split used in the paper's held-out study (deterministic stratified 60/40, seed 42; both non-Quran pairs split 2/2). The evaluation harness (eval.py, research_score) is in the companion code repository (link added upon paper acceptance).

Schema

column	description
id	row id (train-NNN / test-NNN)
transcript	raw ASR transcript (Arabic, no diacritics)
ayah_assignment	gold ayah ids, e.g. 002001-002005; comma-separated segments for non-contiguous recitations; non_quran for negative examples
confidence	reviewer confidence for the gold label
transcript_split_by_ayahs	gold split: transcript text segmented by ayah (`
actual_ayahs	reference ayah ids covered by the recitation

Ayah ids are SSSAAA (3-digit surah + 3-digit ayah), e.g. 095001 = Surah 95 (At-Tin), ayah 1.

quran_ref.json is a slim Quran reference used by the harness: {surah: [{id, ar, clean}]} with ar the Uthmani text and clean a harakat-free canonicalized form (public text).

Splits

split	rows	Quran	non-Quran
train	151	149	2
test	107	105	2

Provenance & privacy

• Recordings were made by consenting users of the application; only ASR-derived text is released — no audio.
• The dataset is de-identified: user and recording identifiers are removed and replaced by sequential row ids. The 4 non-Quran transcripts were manually reviewed and contain no personal content.
• Quranic text itself is public domain.

Versioning

This release corresponds to dataset freeze v1.1 in the paper (the frozen inputs are pinned by SHA-256 in the companion repository).

SHA-256 of released files:

train.csv        fa2cb114df7972946f3f48f0384b5081e94a1a8c8d46a9e3713cb3d9dc9386fa
test.csv         deef397e82f54106eac65e40fcc75b02cd44fbc5b1d03ab3f345842930938fd4
quran_ref.json   0795fe4e3dd01ebdce3f732ba3645226849e299e07a6db7e40eee5c8882baca7

Citation

Citation will be added upon paper acceptance (double-blind review in progress).