diff --git a/LICENSES.md b/LICENSES.md new file mode 100644 index 0000000000000000000000000000000000000000..e7fd32a500383c6341b097d62e3fbe3c0830d870 --- /dev/null +++ b/LICENSES.md @@ -0,0 +1,30 @@ +# Licenses, Attribution & Removal — quranlab/quran-audio-text + +This derived dataset combines components under different terms. The binding +terms remain in the per-row fields. + +## Quran text + +- **Uthmani v1.1** and **Simple-Clean v1.1** +- Source: Tanzil Project, https://tanzil.net/download/ +- License: **CC BY 3.0** +- Both texts are copied verbatim; no normalization or rediacritization is + performed by this dataset. + +## Recitation references + +This dataset hosts **no audio bytes**. Each `audio_url` is copied from +`quranlab/quran-audio` and points to everyayah.com. The recording remains © its +reciter/producer and is included by external reference with attribution. +`commercial_ok=false`; no open redistribution license is asserted. + +Rights holders may request removal of reference/metadata rows through the +dataset's Hugging Face Community (Discussions) tab. + +## Word timing + +When a per-recitation config contains `segments`, the timing fields and their +source/license/attribution are copied unchanged from `quranlab/quran-audio`. +Those timing data are CC BY 4.0; inspect `timing_source`, `timing_license`, and +`timing_attribution` on each row. The derived `timing_quality` label identifies +strict rows and known source-alignment artifacts without changing that payload. diff --git a/QUALITY_REPORT.md b/QUALITY_REPORT.md new file mode 100644 index 0000000000000000000000000000000000000000..f3a1fc9149a12c4f11d06c1f603fe8aaaa0c21e0 --- /dev/null +++ b/QUALITY_REPORT.md @@ -0,0 +1,91 @@ +# Quran audio-text data quality report + +**Status:** PASS +**Validated:** 2026-07-28T13:09:16.217706+00:00 +**Grain:** one row per `(recitation_id, verse_key)` + +## 1. Dataset and grain summary + +| Metric | Value | +|---|---:| +| Canonical ayahs | 6,236 | +| Included strict-Hafs recordings | 43 | +| Expected text-audio pairs | 268,148 | +| Actual manifest rows | 268,148 | +| Unique pair keys | 268,148 | +| Empty Uthmani texts | 0 | +| Empty Simple-Clean texts | 0 | +| Empty audio URLs | 0 | +| Uthmani mismatches | 0 | +| Simple-Clean mismatches | 0 | +| Inherited audio columns with any mismatch | 0 | +| URL-coordinate mismatches | 0 | +| Ayahs with word timing | 268,138 | +| Strict timing rows | 267,579 | +| Flagged source timing artifacts | 559 | +| Missing optional timing rows | 10 | + +## 2. Checks performed + +| Result | Check | Evidence | Severity if failed | +|---|---|---|---| +| PASS | config inventory | 44/44 expected configs; missing=[] extra=[] | none | +| PASS | manifest row count | 268,148 == 268,148 | none | +| PASS | composite-key uniqueness | 268,148 unique (recitation_id, verse_key) pairs | none | +| PASS | manifest schema | 52 ordered columns | none | +| PASS | per-recording spine completeness | 43 recordings, each 6236 rows and unique verse keys | none | +| PASS | surah coverage | 114 == 114 | none | +| PASS | non-Hafs exclusion | excluded IDs present=[] | none | +| PASS | strict reading invariants | riwayah=hafs-asim, enumeration=kufan-6236, basis=hafs-native, risk=false | none | +| PASS | manifest canonical key order | 0 cell mismatches across all repeated spines | none | +| PASS | Uthmani exact equality | 0/268,148 mismatched rows | none | +| PASS | Simple-Clean exact equality | 0/268,148 mismatched rows | none | +| PASS | required payload completeness | empty text=0, text_simple=0, audio_url=0 | none | +| PASS | provenance completeness | text_source=0, text_source_url=0, text_license=0, text_attribution=0, source=0, source_url=0, license=0, attribution=0 | none | +| PASS | source audio column equality | all inherited columns identical across 43 recording configs | none | +| PASS | per-config text equality | 0 transcript mismatches | none | +| PASS | all-row SSSAAA URL mapping | 0/268,148 URL-coordinate mismatches | none | +| PASS | no embedded audio bytes | no `audio` payload column in manifest or per-recitation configs | none | +| PASS | derived timing-quality classification | recomputed mismatches=0, unknown_labels=[], strict=267,579, artifact=559, missing=10 | none | +| PASS | generated timing strict integrity | 199,544/199,544 generated timed rows strict; artifacts=0 | none | +| PASS | inherited cpfair timing artifact bound | 559/68,594 (0.815%) labeled source-artifact; other-source artifacts=0; threshold=<2% | none | +| PASS | manifest/config reconciliation | manifest exactly equals concatenated per-recitation views | none | +| PASS | SHA-256 build provenance | metadata_ok=True, hash_failures=[] | none | +| PASS | upstream filename coverage | 268,148/268,148 paths present for exact_scope=True at 2026-07-28T13:09:16.216974+00:00 | none | + +## 3. Findings + +- No failed local integrity checks. Every output pair is complete, unique, and exactly inherited from its text and audio sources. +- **Low — inherited timing artifacts:** 559/268,138 timed rows are labeled `timing_quality=source-artifact`; 267,579 are `strict` and 10 have no optional segments. The source segments remain verbatim; filter on `timing_quality` for word-level training. + +Upstream existence evidence: 268,148/268,148 referenced filenames present; all_covered=True; checked 2026-07-28T13:09:16.216974+00:00. + +## 4. Temporal or trend anomalies + +The Quran spine is static, so temporal drift is not applicable to the text. +External audio availability can change; the timestamped checksum inventory +above is the freshness evidence for the current build. + +## 5. Likely causes and impacted use cases + +If a future check fails, the most likely causes are a rebuilt source Parquet, +an audio registry scope change, an upstream filename removal, or an accidental +non-Hafs join. Any such failure blocks ASR/TTS use because it can create a +wrong transcript-label pair. + +## 6. Recommended fixes and automated tests + +- Keep this validator as a release gate. +- Refresh `--verify-upstream` before every publication. +- Never admit a non-Hafs recording without a verified matching text spine. +- Treat missing upstream files as failed rows; do not silently retain them. + +## 7. Assumptions and open questions + +- File existence and coordinate mapping are verified without downloading + copyrighted audio bytes. Semantic listening of all recordings is outside + this reference-only dataset's legal/operational scope. +- `text` and `text_simple` are exact source strings, not normalized variants. +- Word timing is inherited and separately attributed; verse pairing does not + require every row to have complete word timing. Use + `timing_quality == "strict"` when exact word-level alignment is required. diff --git a/README.md b/README.md new file mode 100644 index 0000000000000000000000000000000000000000..fa450e94fbfd190f3266a8539571b6b16636117c --- /dev/null +++ b/README.md @@ -0,0 +1,358 @@ +--- +license: other +license_name: mixed-cc-by-text-timing-and-reference-only-audio +pretty_name: "QuranLab — Verse-Aligned Quran Text + Recitation References" +language: + - ar +task_categories: + - automatic-speech-recognition + - text-to-speech +tags: + - quran + - islam + - arabic + - audio + - text + - recitation + - automatic-speech-recognition + - text-to-speech + - verse-aligned + - hafs + - reference-dataset +size_categories: + - 100K