A personal extension project by Raufa, with 50 requested reciters selected by Quran Lab. Batch 1 covers list entries 1–30; batch 2 covers 31–50. Source coverage follows the selected MP3Quran Hafs recording and may be incomplete. Entry 49, “Saleh Al Sadeiq”, remains unconfirmed; the nearest current MP3Quran identity is Salman Alsadeiq (285).

Audio and processing

MP3Quran source → float32 44.1 kHz decode → Anvuew Mel-Band RoFormer dereverb → Quran Lab Zipformer v3.1 CTC inference → known-reference forced alignment → waveform boundary refinement → 24 kHz mono, 24-bit FLAC.

The primary Zipformer candidate uses PyTorch CUDA W8A8 quantization of encoder Linear layers. Unsupported operations remain floating point, and the CTC head/log-softmax remain FP32. The runtime chooses the fastest configuration that passes the FP32 alignment comparison; INT8 performance and quality are not yet measured. Anvuew uses BF16 eager B8, 352800-sample chunks and overlap 2, with overlap-add reconstruction.

The measured 53.134925× source throughput is an Anvuew inference benchmark on A100-SXM4-80GB. It is not a measurement of the entire dataset pipeline. Runtime records separate stages and session CU at 6.7 CU/hour.

Layout and streaming

audio/<qari>/part1.tar                 # up to 2–3 tar per qari
manifests/<qari>.jsonl
sources/<qari>.json
shards/<qari>.json
progress/<qari>.json
qc_flags/<qari>.jsonl
metadata.jsonl
dataset_stats.json
notebooks/QuranTTS_Extended_A100.ipynb
pipeline/QuranTTS_Extended_bundle.zip

Each tar contains pairs such as 001_001.flac and 001_001.json. Manifests reference the archive using shard_path and member_path; file_name is a logical audio path. Reading a tar-backed row requires the supplied streaming loader; the logical path is not a separately uploaded FLAC file.

The notebook runs all 30 batch-1 reciters in one pass. A single live text panel tracks the active reciter, current surah, completed cleaning/alignment surahs and step-based percentages. Detailed logs remain in a file. Four-ayah previews are saved only to mounted Google Drive, under MyDrive/QuranTTS_Extended/samples/batch1/, as a listening MP3, FLAC master and ayah index JSON.

Metadata follows the QuranTTS fields: reciter, source_id, riwaya, style, surah, ayah, text_uthmani, text_imlaei, phoneme_tokens, n_tokens, duration_s, align_score, sample_rate, bit_depth, src_part, src_start_s, src_end_s and boundary. Extensions include phonetic context, token timing, alignment precision, checksums and pinned model/source revisions.

Labels and alignment

Canonical labels come from Quran Lab quran-tajweed-phonetics, revision 6f4be6257c140f7e318574bca01386474b606999. Waqf/wasl context uses Quran Lab quran-g2p, revision c5f9125951c75c2a7f82119fac9f272f1df62df6.

The acoustic model is Quran Lab Zipformer v3.1, revision 3da514fc833b22c902ee466a7710cf399896f2f0. Original model blank=250 is normalized to alignment blank=0 using the same rotation for target IDs and posterior columns. Inference uses cache-aware 61-frame input chunks, advancing 48 feature frames, with a nominal 40 ms CTC output grid.

Greedy decoding supplies anchors; CTC Viterbi supplies the complete reference path. Acoustic quiet intervals refine clip boundaries. Blank probabilities do not define silence. Madd and ghunnah are preserved by retaining the complete ending, rather than cutting at the last CTC peak. Timing remains an automatic estimate, not manually established phoneme duration.

Isti‘adzah and unnumbered introductory basmalah are excluded from numbered ayat. Al-Fatihah 1:1 retains basmalah. Repetitions, incomplete readings, uncertain boundaries and failed alignment are recorded; the first complete supported occurrence is preferred. Four accepted ayat per reciter form a deterministic hearing preview.

Access, provenance and licensing

The repository is public with manual gated access. Gating controls downloads; it does not change underlying rights. Phonetic labels and G2P retain Quran Lab NPL-1.2 attribution and terms; see LICENSE-PHONETICS and the vendored G2P license. Pipeline code is MIT; vendored icefall is Apache-2.0. Source recording rights remain with their respective holders, and are not relicensed by the phonetic-label license. The source manifest identifies the recording and URL for every processed surah.

QuranTTS supplies the metadata-format reference. Its original audio and metadata rows are not copied into this extension. Text editions are attributed to Tanzil, with the text/license notices retained in the distribution.

Automatic alignment and phonetic labels may contain errors. They are training annotations and do not establish a religious judgment on recitation.

Current statistics

Actual counts are published in dataset_stats.json and updated after each committed qari. Until processing begins, accepted clips, source hours, output hours and measured pipeline CU are all zero. The 50-reciter target is not a claim of completed coverage.

Downloads last month: 6

Total file size:

817 kB