QuranTTS

An ear-verified Quranic recitation corpus for speech restoration and TTS. Three configurations ship in this repository: the same ayah-aligned segments in cleaned and uncleaned form, plus the larger raw pause-cut pool they were cut from.

from datasets import load_dataset

# ayah-aligned, cleaned, with text and phonemes  (default)
ds = load_dataset("Quran-Lab/QuranTTS", "v2_clean", split="train")

# the SAME segments and labels, before any cleaning
raw = load_dataset("Quran-Lab/QuranTTS", "v2_raw", split="train")

# the larger raw pause-cut pool, audio and source labels only
pool = load_dataset("Quran-Lab/QuranTTS", "v1_chunks", split="train")

Which configuration do I want?

	v2_clean	v2_raw	v1_chunks
Segments	1,615	1,615	14,091
Duration	7.7 h	7.7 h	58.4 h
Segment unit	one complete ayah	one complete ayah	pause-to-pause chunk
Arabic text	yes	yes	no
Phonemes	yes	yes	no
Surah and ayah ids	yes	yes	no
Acoustic cleaning	yes	no	no
Sample rate	48 kHz mono FLAC	48 kHz mono FLAC	48 kHz mono FLAC

v2_raw and v2_clean are matched pairs: identical segments, identical boundaries, identical labels, differing only in whether the cleaning chain was applied. That gives directly usable (degraded, clean) training pairs for speech restoration, and it lets you reject our cleaning choices and redo them from the aligned audio.

Use v2_clean for anything needing labels and clean targets. Use v2_raw when you want the original acoustics or your own processing. Use v1_chunks when you want maximum audio and intend to do your own segmentation.

On the size difference

v1_chunks holds 58.4 h, of which the ayah-aligned configs currently cover only 7.7 h. That gap is a limitation of the alignment stage, not a judgement about the remaining audio. The locator works on ten-minute spans and discards a span it cannot confidently place, and it resolves each span to a single surah, so recordings that cross surah boundaries lose material. Sources that sit inside one surah converted at 60 to 77 percent; the longest continuous recitations converted at under 5 percent. The unconverted audio is all present in v1_chunks. Improving this is the main open work on the dataset.

v2_clean and v2_raw

Fields

field	description
audio	48 kHz mono FLAC, one complete ayah
source_id	originating recording
surah, ayah	canonical location, 1-indexed
text_uthmani	Uthmani script, as recited
text_imlaei	imlaei (simplified) orthography
phonemes	Quran Phonetic Script, whole-ayah phonetisation
duration_s	seconds
reciter, grade, tier	source-level provenance
cleaning	the exact processing chain applied

Phonemes are produced with quran-transcript's Hafs phonetiser, applied to the whole ayah rather than word by word. Per-word phonetisation measures 20.5% PER against whole-ayah 1.6%, because cross-word tajweed rules (idgham, madd at word boundaries) are invisible when words are processed in isolation.

How the audio was segmented

Rather than cutting on silence and hoping, each recording is transcribed with a CTC model, the decoded text is located inside the canonical Quran with a sequence matcher over a normalised 6,236-ayah index, and only the ayat actually present are force-aligned. Boundaries are then refined against signal onsets. Spans that locate nothing, such as introductions or nasheed beds, are dropped rather than guessed at.

Cleaning chain

de-hum notches (50/60 Hz and harmonics, Q=30)
  -> iZotope RX Dialogue Isolate (dialogue 0, reverb -30, noise -20)
  -> dereverb-echo mel-band roformer
  -> iZotope RX Voice De-noise (reduction 6)
  -> high-pass 80 Hz
  -> resample to 48 kHz
  -> normalise to -1 dBFS true peak

Every stage is subtractive. No generative model touches this audio, so nothing can be hallucinated into a recitation. That was a deliberate constraint: a generative enhancer can smooth or lengthen a vowel, and in Quranic recitation an altered vowel length is an altered madd.

Measured effect across the corpus:

	before	after
noise floor	-18.2 dB	-21.3 dB
4-8 kHz energy share	0.234 %	0.299 %
spectral edge	17.9 kHz	19.1 kHz

The 4-8 kHz band rising matters more than the floor: that region carries sibilants and the emphatic consonants, and most denoisers erode it. The band edge rising rather than falling is why a second, more aggressive dereverb model was rejected during development, since it silently brickwalled everything above 17.5 kHz.

All 1,615 files were verified after processing for sample rate, clipping, dynamics, and duration drift against their source segment, so the text and phoneme labels still align with the audio.

Known limitations

• Hafs only. Other qira'at are not represented.
• Small speaker pool. This is a clean-target corpus, not a speaker-diversity corpus. v2_clean draws on 21 recordings.
• Coverage is partial, 923 unique ayat across 19 surahs. It is not a complete mushaf.
• Cleaning is not uniform. Roughly 7% of segments came out with a slightly worse noise floor than they started with, scattered across sources rather than concentrated in any one.
• Waqf (pause) rules are not modelled in the phoneme layer.
• v1_chunks segments are cut at energy minima, which occasionally fall mid-breath. Do not treat those boundaries as phrase boundaries.

Provenance and licensing

Audio originates from publicly posted recitations, each auditioned and approved by a human listener before inclusion. Sources that measured well but sounded processed were rejected. Full per-source provenance is maintained offline by the maintainers.

Released under NPL 1.1 (see LICENSE). If you believe a recording of yours is included and you want it removed, open a discussion on this repository and it will be taken down.