Mathani-Ayat/qaloon-reciter-dataset

Qaloon Reciter — continuously maintained data

Initial release: 0.1.0; schema 1.0.0. A private team working corpus for Qālūn ʿan Nāfiʿ ASR. Current scope: al-Fātiḥah + Juzʾ ʿAmma (78–114), not the whole Quran. Three reciters have been approved by the project owner for this team release: Huthaify, Husary and Dokali. This approval is not a claim of a verified source copyright license or acoustic alignment review.

Included and excluded material

• 569 recordings per approved reciter; 1,707 WAVs in the inventory.
• Format: mono 16 kHz, signed 16-bit PCM WAV, no MP3 re-encoding of slices.
• The >30 s Husary clips remain in the inventory but are excluded from the ≤30 s ASR split manifests; no truncation against full-ayah labels.
• sources.json records Waleed and Trabulsi as unauthorized. Their audio and clip-level metadata are not uploaded. A small Trabulsi audit summary records structural gaps and candidates only.
• Saltany is outside this release and is untouched.
• Dokali has known boundary-bleed risk: boundary_review_recommended is attached to his clips. These remain weakly aligned, not gold word-level data.

Structure and use

dataset_qaloon_hutafi/metadata.jsonl
dataset_qaloon_hutafi/audio/078001.wav
dataset_qaloon_hutafi/audio/metadata.jsonl
dataset_qaloon_Husary/...
dataset_qaloon_dokali/...
train.jsonl, validation.jsonl, test.jsonl
release_manifest.json, sources.json, excluded.json

Recommended training path: snapshot_download the pinned dataset revision, then use the project loader on the downloaded folder with explicit --include-reciter huthaify husary dokali. It retains the reciter-directory structure and verifies WAV durations. The root split manifests also have repo-relative audio paths and can be resolved against the snapshot root. The three AudioFolder configurations are convenient per-reciter inventory views; their train name means inventory, not the leakage-safe experimental training partition. Use the split field or the root split manifests for experiments, never treat all inventory rows as train.

from huggingface_hub import snapshot_download

root = snapshot_download("BadrSh/qaloon-reciter-dataset", repo_type="dataset",
                         revision="<PINNED_COMMIT_FROM_RELEASE_MANIFEST>")

Schema

Rows preserve the original transcript fields and source annotation IDs:

Field	Meaning
sample_id	Stable reciter-surah-ayah ID; filenames alone collide across voices
surah, ayah	Corrected Qālūn text IDs; not necessarily raw source numbering
source_ayahs	Original Uthmani source annotation IDs contributing to the text
text_raw_uthmani	Source text with original annotations/numerals
text_asr_normalized	Unvowelled Qālūn-aware ASR label; preserves medial ئ/ؤ and terminal ى
normalized_with_harakat	Cleaned source vowels; not a tajweed-error annotation
relative_audio_path	Relative to reciter directory in its metadata; relative to snapshot in root split files
duration_seconds, audio_sha256	Verified WAV duration and content identity
source_url, mp3quran_read_id	Public source location; signed URLs and credentials excluded
authorization_status, source_license	Team inclusion status and source-license evidence
alignment_status, quality_flags	Explicit quality status rather than an implied gold label
split, usable_for_asr	Stable partition and ≤30 s eligibility

Evaluation discipline

Seed 42, SHA-256 ayah-based split, approximately 80/10/10. All recordings of an ayah share a partition; Fātiḥah is grouped as one unit because source numbering and basmalah conventions differ. release_manifest.json stores the actual counts. Voices appear in multiple partitions: this tests unseen ayahs, not unseen speakers. Omission/tajweed precision requires separately labelled real learner errors. WER/CER alone do not certify a teacher replacement.

Updates and reproducibility

Each update is a Hub commit plus dataset_version and schema_version. Keep sample IDs stable; pin a commit for experiments, not the moving main branch. Audio checksums expose replaced clips even when filenames stay fixed. New sources need provenance, authorization, verified transcript boundaries, duplicate checking and compatible original-source numbering. Never guess missing Trabulsi ayahs from gaps or concatenate clips based on names alone. Fallback voices are separate samples with their true reciter identity, never pretend a replacement clip is Trabulsi.

Provenance, limitations and terms

MP3Quran audio is paired with the project's Qālūn Uthmani JSON; the source text edition and redistribution permission still need bibliographic/license review. No CC/MIT license is invented for recordings we did not create. The initial repository is private while provenance is documented. See LICENSE and sources.json before redistribution. A later public release should contain only material with demonstrable compatible redistribution rights.

Maintainer: Ai-Reciter project.

Additional Garu configuration

Qaru's segmented audio is available under dataset_qaloon_garu/: 547 labelled clips and 16 review-only clips. See that folder's README and release manifest. The garu config is an inventory view; unresolved surah 91 clips have pending_numbering_review=true and usable_for_asr=false. They are excluded from the labelled metadata.jsonl. The original root split manifests are still the three-reciter experiment, not a pooled Garu training split. Source rights/alignment are not certified by this upload.