ar-quran-cm17-timestamps

The timestamp-supervised variant of the resource. Same audio pipeline and text conventions as ar-quran-hadith14books-MSA, but the transcripts carry Whisper-style time tokens, so a fine-tune can keep predicting timestamps instead of losing that ability.

Why it exists. Naively fine-tuning Whisper on untimestamped text erases its timestamp prediction entirely — catastrophic forgetting of the timestamp task. Mixing this data into 23.4 % of the training rows is what preserves it, and that is how whisper-large-v3-ar-eg-timestamps was produced. Timestamps matter here for a specific reason: they let a reader jump back to the exact verse in the recording and verify it, which is the whole point when the text is scripture.

• Paper: A Quran and Hadith Speech Resource and Benchmark for Arabic ASR (preprint, DOI) · Code: github.com/DrAliGomaa/quran-hadith-asr
• Created by Mohamed Kotb · Under the patronage of Prof. Ali Gomaa, former Grand Mufti of Egypt, who reviewed the handling of the religious material.

Splits

Split	Rows	Content
train	2,887	Quran (reciter al-Ḥuṣrī) + Common Voice 17, with word/segment timestamps
cm_17_validated	7,957	Common Voice 17 validated Arabic
eg	5,034	Egyptian (Prof. Ali Gomaa)
quran_hadith	7,344	Quran + Hadith

Transcript format

Timestamped rows use Whisper's time-token convention — a start token before each segment and an end token after, quantized to 20 ms:

<|0.00|> بسم الله الرحمن الرحيم <|3.28|><|3.28|> الحمد لله رب العالمين <|6.12|>

Untimestamped rows are plain text. A predict_timestamps flag routes each example to the timestamped or non-timestamped decoder prefix during training, which is what lets both live in one mix.

Load it

from datasets import load_dataset, Audio

ds = load_dataset("Dr-AliGomaa/ar-quran-cm17-timestamps", split="train")
ds = ds.cast_column("audio", Audio(sampling_rate=16000))
print(ds[0]["sentence"])     # contains <|start|> ... <|end|> tokens

Strip the time tokens if you want plain text:

import re
plain = re.sub(r"<\|[\d.]+\|>", " ", ds[0]["sentence"]).strip()

Usage notes

• Apply the same audio pipeline at inference as the main dataset — see its card. The timestamps were computed on pipeline-processed audio, so raw audio shifts them.
• Text conventions are identical: Imlāʾī Quran orthography, no tashkīl, numerals as words.
• Licensing follows the component sources listed on the main dataset card.

Citation

@misc{kotb2026quranhadith,
  title  = {A Quran and Hadith Speech Resource and Benchmark for Arabic ASR,
            with Professional-Reciter Training and Validation},
  author = {Mohamed Kotb},
  year   = {2026},
  publisher = {Zenodo},
  doi    = {10.5281/zenodo.21927416},
  url    = {https://doi.org/10.5281/zenodo.21927416},
  note   = {Preprint}
}