Bayanat Labs Open Arabic Speech Hours Counter

Arabic speech datasets Arabic speech hours speech dataset licences Arabic ASR data open speech data

Key numbers

Counted by script from the released files. Each key number is read from feed/latest.json, qa/second_check_summary.json or evidence/measurements/yodas3-ar.json in the Hugging Face dataset and pinned to its SHA-256; MANIFEST.json lists every file with its size and SHA-256.

Downloads and identifiers

Get started

Load the table with the datasets library, or embed the counter on a page: the widget makes one request for latest.json, sends no cookies, stores nothing and shows the as-of date. To re-run the measurements, clone the repository at the v0.1.0 tag; the package is not on PyPI yet.

# Python
from datasets import load_dataset
rows = load_dataset("Bayanat-Labs/arabic-speech-hours", "corpora", split="train")

# Embed the counter in a web page
<arabic-speech-hours src="https://huggingface.co/datasets/Bayanat-Labs/arabic-speech-hours/resolve/main/feed/latest.json"></arabic-speech-hours>
<script src="https://huggingface.co/datasets/Bayanat-Labs/arabic-speech-hours/resolve/main/embed/speech-hours.js" defer></script>

# Re-run the measurements
git clone --branch v0.1.0 https://github.com/Bayanat-Labs/arabic-speech-hours.git
cd arabic-speech-hours
python -m pip install .
python scripts/measure.py
python scripts/build.py --as-of 2026-10-30

What is released

Attribution: Open Arabic Speech Hours Counter, Bayanat Labs, CC BY 4.0, as of 30 October 2026. https://huggingface.co/datasets/Bayanat-Labs/arabic-speech-hours

Intended uses

• See how many hours of released Arabic speech audio each licence class holds, and which corpora are behind each figure, before planning speech recognition or synthesis work.
• Check, for one corpus, the licence as its publisher states it, how its hours were obtained and what overlap was removed.
• Show the current figure on your own page with the widget or the badge, which carry the as-of date.

Out of scope

• Legal clearance: licence classes are information, not legal advice; read each publisher's terms before you use a corpus.
• Judging quality, dialect coverage or speaker diversity: the counter counts released hours, not how useful they are for a task.

Rights at a glance

Training, display, redistribution and synthetic-voice use are separate permissions.

Limitations

• The figures are as of 30 October 2026 and change as corpora are released, updated or withdrawn; each row records when its licence page was read.
• Hours come from published metadata or, where none exists, from a figure the publisher states; stated figures were quoted, not measured.
• Overlap is removed only where publishers state it or published video IDs show it, and only within one licence class. Overlap with YODAS could not be measured, because its recording IDs are publisher IDs.
• Overlap that could not be measured is not removed: MASC with YODAS v3 and YODAS2 (open, commercial use allowed) and QASR with MGB-2 (research use or on request), so those class figures may count shared hours twice.
• YODAS v3 files its Arabic subset by caption-track language; the counter does not check the spoken language of the audio, and dialect is recorded only where the source states it.
• Each corpus has one licence class for its audio, read from the publisher's own page; copies that declare another licence are noted, not used. MASC's CC BY 4.0 is read from the dataset metadata of the publisher's IEEE DataPort page, whose visible text states no licence; a third-party copy of MASC on the Hugging Face Hub states CC BY-NC 4.0.
• QASR's card states CC BY-NC 2.0, holds no files and links to a page where the data is obtained through a request process, so QASR is counted as research use or on request.
• Gated repositories were not opened; only their public cards were read. Datasets not yet reviewed are listed as not counted, with the reason.

Sources and licences

Each corpus's licence class comes from the publisher's own page, saved with its SHA-256 hash and retrieval time. Hours come from published metadata read without downloading audio (duration columns of Parquet metadata read with range requests, duration manifests, Common Voice release statistics and the Hugging Face dataset viewer's duration statistics) or, where none exists, from a figure the publisher states, quoted with its location. Requests carried the Bayanat Labs open-data user agent and were spaced at least a second apart per host; gated repositories were not opened. Source details

How to cite

Cite as:

Bayanat Labs (2026). Open Arabic Speech Hours Counter, version 0.1.0, as of 30 October 2026. Bayanat Labs. https://huggingface.co/datasets/Bayanat-Labs/arabic-speech-hours

• CITATION.cff

No DOI. Cite the URL in the citation above and the date you read it.

@misc{bayanatlabs2026arabicspeechhours,
  author       = {{Bayanat Labs}},
  title        = {Open Arabic Speech Hours Counter, version 0.1.0, as of 30 October 2026},
  year         = {2026},
  month        = oct,
  howpublished = {\url{https://huggingface.co/datasets/Bayanat-Labs/arabic-speech-hours}}
}

How we handle citations

Changelog

Corrections and removal requests

If something in this release should not be there, or is wrong, email hello@bayanatlabs.com. We handle removal requests within 30 days, and platform requests within 24 hours. How removal requests work

Related product: Bayanat Voice