QuranTTS

An ear-verified, studio-clean Quranic recitation corpus. 58.4 hours of Hafs recitation in 14,091 pause-aligned segments at 48 kHz mono FLAC, curated for training speech restoration and text-to-speech models where the targets must be acoustically dry: no room reverb, no noise floor, no broadcast processing.

Why this corpus exists

Publicly available Quranic audio is almost universally recorded or mastered with room acoustics: mosque reverb, studio reverb added by convention, or heavy post-processing with audible residue. For listening that is fine. For training restoration models (the clean target side) and TTS vocoders it is poison: a model trained toward reverberant targets learns to produce reverb.

QuranTTS is the result of triple-gating a much larger candidate pool:

1. Signal metrics: 20 dB decay time, noise floor relative to speech, spectral edge, measured per source.
2. Tiering: sources graded for bandwidth and floor depth; grades route each source to the tasks it can serve (fullband sources can train a 48 kHz vocoder; clean-but-narrowband sources still fully serve 16 kHz feature-space training).
3. Human ear: every source was auditioned. The human verdict overrides the meters in both directions; sources that measured excellent but sounded processed were rejected. Fewer than half of the candidate sources survived.

Composition

Total audio	58.4 h in 14,091 segments
Fullband subset (>= 16 kHz true bandwidth)	24.1 h in 5,801 segments
Sources	31 ear-verified recitation recordings
Format	48 kHz mono FLAC, segments 4-18 s cut at pause points
Riwaya	Hafs 'an 'Asim

Files

• data/<source_id>/*.flac - the audio segments
• metadata_chunks.jsonl - one row per segment: file, source, duration, band grade, tier, verification status

Filter on grade == "fullband" when selecting 48 kHz vocoder targets; narrowband segments serve 16 kHz feature-space training only.

Intended use and limitations

• Built as clean-target data for speech restoration (parametric resynthesis) and TTS vocoder training.
• Segments are cut at energy minima, which occasionally fall mid-breath; segment boundaries should not be treated as phrase boundaries.
• The corpus is Hafs-only and skews toward a small number of reciters; it is a clean-target corpus, not a speaker-diversity corpus.

Provenance and licensing

Audio originates from publicly posted recitations. Full per-source provenance is maintained offline by the maintainers; if you believe a recording of yours is included and want it removed, open a discussion on this repository and it will be traced and taken down.

Released under the Quran-Lab No-Profit License v1.1 (see LICENSE): free for non-profit use in service of the Quran; commercial use prohibited.

Curation

Curated by the Quran-Lab effort as part of a Quranic speech-restoration and TTS pipeline. Companion resources: Quran-Lab/mfa-quran-hafs (forced alignment for tajweed measurement).