Luxbg/quran-everyayah-aligned

EveryAyah with word timings

The training, validation and test data used for Luxbg/quran-turbo-streaming-lora, a causal streaming version of a Quran Whisper model. It is EveryAyah (tarteel-ai/everyayah) restricted to clips of at most 30 seconds, with a start and end time for every word, produced by CTC forced alignment.

Split	Clips	Hours	Selection
Train	156,075	561	every clip up to 30 s that aligned (167,908 candidates)
Validation	1,932	7	every 10th clip up to 30 s that aligned
Test	1,963	7	every 10th clip up to 30 s that aligned

The original EveryAyah splits are kept. All three share reciters.

Contents

train/<reciter>.tar     one archive per reciter collection (36), audio and word timings
validation.tar, test.tar
meta/train.tsv, meta/validation.tsv, meta/test.tsv     manifests
meta/train-manifests.tar                              per-shard manifests from preparation

Each clip is a 16 kHz mono FLAC (lossless, sample-identical to the source WAV) with a .wrd file beside it, one line per word: start end word, with start and end in samples. Manifest columns: wav_path, tg_path (the .wrd file), raw_text (Tanzil "simple" text with diacritics, as in EveryAyah), reciter, duration (seconds), align_score (mean log-probability per frame of the alignment path).

Extract the archives at the repository root. The manifests hold absolute paths from the machine that built them (/workspace/careless-whisper-quran/prepared/...); replace that prefix with the extraction directory.

How the timings were made

Alignment used a FastConformer CTC model trained on Quranic recitation (Muno459/fastconformer-quran, ONNX, a copy of which is in the model repository under extras/). Each clip is matched to its ayah by bare letters and aligned against the assimilated spelling that model was trained on (99.95% of clips matched); the timings are then mapped back to EveryAyah's words one to one. Short vowels written before shadda are reordered to match the aligner's tokenizer. Because the aligner places the final token on the last frame of every clip, the end of the last word is taken from audio energy instead: the last 80 ms frame above 10% of the clip's peak RMS. A word ends where its last emitted token ends, so pauses are not counted as part of the preceding word.

A clip was skipped (7.0% of training candidates) when its text matched no ayah, when alignment failed, or when the alignment's mean log-probability per frame was below −1.0.

Known issues

• Some recordings contain a phrase twice because the reciter restarted. The text has it once, so the timings for those clips were aligned to text without the repeat. About 2% of clips, estimated from the test set.
• Some reciters' source audio is strongly band-limited (little or no energy above 3 to 4 kHz), consistent with a low-bitrate original.
• Transcripts of the longest ayat exceed Whisper's 448-token limit once every letter and diacritic is tokenized.

License and source

Derived from Tarteel's EveryAyah dataset; see its card for the terms of the audio and text. Prepared for non-commercial research, under the same CC BY-NC 4.0 terms as the model it was built for. Code: quran/prepare_data.py and quran/align.py in the careless-whisper-quran repository.