Bayanat-Labs/arabic-speech-hours

Open Arabic Speech Hours Counter

How many hours of Arabic speech have been released, and what may you do with them? This table counts released Arabic speech hours by licence class, hour type, dialect and genre, from published metadata and quoted publisher statements, for 17 corpora. Version 0.1.0, as of 30 October 2026. By Bayanat Labs.

As of 30 October 2026, 36,221 hours of released Arabic speech audio carry an open licence that allows commercial use, across 8 corpora, after removing overlap that the publishers state. 99.5% of these hours come from corpora derived from YouTube videos (as their publishers state), and 3.0 hours (0.008%) carry a dialect label other than Modern Standard Arabic from the source. YODAS v3 accounts for 92.4% of the figure; its Arabic subset is filed by caption-track language. By uploader locale, 6,898.4 of its 33,460.7 hours are Arabic, 16,310.0 another language and 10,252.3 not given.

Beside it, never added to it: open, non-commercial only 667 h (1 corpus); no derivatives 16 h (1 corpus); research use or on request 3,216 h (4 corpora); no licence found 4,639 h (3 corpora).

See the chart and table in the browser: Bayanat-Labs/arabic-speech-hours (English and Arabic, works offline).

How to use

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

The latest figures are in feed/latest.json (schema: feed/latest.schema.json), and a badge in feed/badge.svg. To embed the counter on a 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>

The widget (embed/speech-hours.js, 10 KB) makes one request (latest.json), sends no cookies, stores nothing, shows the as-of date and links to https://www.bayanatlabs.com/open/arabic-speech-hours. It follows the language and direction of its container (lang="ar" for Arabic).

Released audio by licence class

Licence class	Released audio hours (overlap-adjusted)	Gross	Corpora
Open, commercial use allowed	36,221.0	36,223.7	8
Open, non-commercial only	667.0	667.0	1
No derivatives	15.9	15.9	1
Research use or on request	3,216.3	3,216.3	4
No licence found	4,639.1	7,672.1	3

Hour types are kept apart: transcribed hours (open, commercial use allowed 907 h; open, non-commercial only 438 h) and Common Voice validated hours (93.3 h) are separate rows and are not part of these totals.

Corpora

Corpus	Licence as stated	Class	Released audio, h	Overlap-adjusted, h	How the hours were obtained
Common Voice Scripted Speech 27.0, Arabic	CC0-1.0	Open, commercial use allowed	158.3	158.3	computed from release statistics
YODAS v3, Arabic subset	CC BY 3.0	Open, commercial use allowed	33,460.7	33,460.7	computed from metadata
YODAS2, Arabic subset (ar000)	CC BY 3.0	Open, commercial use allowed	1,569.4	1,569.4	computed from metadata
MASC: Massive Arabic Speech Corpus	CC BY 4.0	Open, commercial use allowed	1,000.0	1,000.0	stated on the card or in a paper
Arabic Speech Corpus	CC BY 4.0	Open, commercial use allowed	3.0	3.0	stated on the card or in a paper
ArVoice	CC BY 4.0 (repository); Parts 1 and 2 by data use agreement	Open, commercial use allowed	2.7	0.0	stated on the card or in a paper
ArVoice	CC BY 4.0 (repository); Parts 1 and 2 by data use agreement	Research use or on request	7.3	7.3	stated on the card or in a paper
ClArTTS	CC BY 4.0	Open, commercial use allowed	11.4	11.4	computed from metadata
FLEURS, Arabic (ar_eg)	CC BY 4.0	Open, commercial use allowed	8.2	8.2	computed from metadata
MediaSpeech, Arabic	CC BY 4.0	Open, commercial use allowed	10.0	10.0	computed from metadata
ADI17	none found	No licence found	3,078.3	3,078.3	stated on the card or in a paper, computed from metadata
ADI-20 (additions to ADI17)	none found	No licence found	3,556.8	523.8	stated on the card or in a paper
SADA: Saudi Audio Dataset for Arabic	CC BY-NC-SA 4.0	Open, non-commercial only	667.0	667.0	stated on the card or in a paper
QASR	CC BY-NC 2.0 (card metadata); "Non-Commercial Purpose ONLY!" (card text)	Research use or on request	2,000.0	2,000.0	stated on the card or in a paper
MGB-2	QCRI-ALT Corpus License Agreement (non-exclusive research use)	Research use or on request	1,200.0	1,200.0	stated on the card or in a paper
Casablanca (validation and test release)	CC BY-NC-ND 4.0	No derivatives	15.9	15.9	computed from metadata
Mixat	research use (paper); no licence on the card	Research use or on request	9.0	9.0	computed from metadata
Tarteel AI Quran verse recitations (tarteel-ai/everyayah)	recordings not documented; remainder CC BY 4.0	No licence found	1,037.0	1,037.0	computed from metadata

The corpora configuration has one row per corpus, partition and hour type (31 rows), with the licence string and quote, the SHA-256 hash and retrieval time of the licence page, access, dialect group and the source's own dialect label, genre, seconds, hours, the method and evidence for the hours, overlap-adjusted hours, and the second-check result.

What is counted

• Hours of Arabic speech audio that a publisher has released. Hours are stored in seconds and rounded only for display.
• Hour types: released audio (the headline basis), transcribed, validated (Common Voice), clean and manifest. They are never added together.
• Licence classes: open, commercial use allowed (CC0, CC BY, CC BY-SA, ODC-By, public domain, or a software licence declared for data); open, non-commercial only (CC BY-NC variants); no derivatives (any ND variant); research use or on request (research licences, data use agreements, request forms or request processes); manifests only (identifiers or timestamps without audio); disputed (the publisher states two different licences); no licence found. A corpus has one class for its audio; the classes are shown side by side and never pooled.
• The licence is read from the publisher's own page. Copies that declare another licence are noted in the row's notes (for example, one MASC copy declares CC BY-NC 4.0 and one SADA copy declares CC BY 4.0); the publisher's licence is used.
• Dialect is recorded as the source states it, with the source's label kept; a country label is marked "source-stated country". Without a statement the group is "unlabelled". Quran recitation is a separate genre.
• Bayanat Labs releases are counted only for released recorded speech. The Bayanat Labs channel-pairs set renders FLEURS recordings through simulated channels; it is derived audio and is listed as excluded.

Licence notes

• MASC: the licence (CC BY 4.0) is read from the dataset metadata of the publisher's IEEE DataPort page; the visible page text states no licence. A third-party copy on the Hugging Face Hub (abdusah/masc) states CC BY-NC 4.0. The counter uses the publisher's licence, so MASC is in the class “Open, commercial use allowed”.
• QASR: the publisher'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 in the class “Research use or on request”.

How the hours are obtained

• Computed from metadata: duration columns of Parquet metadata read with HTTP range requests (for YODAS v3, 20.4 MB were read from 208 metadata files of 1.6 GB), per-recording duration manifests (YODAS2), sample counts in manifests (FLEURS), segment tables (MASC transcribed hours, ADI17 validation and test segments), or the Hugging Face dataset viewer's duration statistics (ClArTTS, Casablanca, MediaSpeech).
• Computed from release statistics: Common Voice per-locale statistics from the Common Voice release statistics repository.
• Stated on the card or in a paper: a quoted figure from the publisher's page or a paper describing the corpus, with its location.
• The counter does not download audio. Gated repositories are not opened; only their public cards are read.

Overlap

Hours that two corpora share are removed once, from the later corpus in the priority order, and only when both are in the same licence class.

• YODAS v3, Arabic subset and YODAS2, Arabic subset (ar000): the publisher states that the two hold different recordings; nothing removed.
• Arabic Speech Corpus and ArVoice (part-3): publisher statement; 2.69 h shared, 2.69 h removed from ArVoice.
• ADI17 and ADI-20 (additions to ADI17): publisher statement; 3,033.00 h shared, 3,033.00 h removed from ADI-20 (additions to ADI17).
• MASC: Massive Arabic Speech Corpus and ADI17: published video IDs; 3 of 1,375 ADI17 video IDs in scope also appear among 12,917 MASC: Massive Arabic Speech Corpus video IDs (0.01 h of ADI17 segments); scope: MASC segment tables (all subsets) and ADI17 validation and test segment IDs; reported only, because the two are in different licence classes.
• MASC: Massive Arabic Speech Corpus and ADI-20 (additions to ADI17): published video IDs; 107 of 59,644 ADI-20 (additions to ADI17) video IDs in scope also appear among 12,917 MASC: Massive Arabic Speech Corpus video IDs (3.43 h of ADI-20 (additions to ADI17) segments); scope: MASC segment tables (all subsets) and the ADI-20-53h training manifest; reported only, because the two are in different licence classes.
• MASC: Massive Arabic Speech Corpus and YODAS v3, Arabic subset: not measured: YODAS recording IDs are publisher IDs.
• MASC: Massive Arabic Speech Corpus and YODAS2, Arabic subset (ar000): not measured: YODAS recording IDs are publisher IDs.
• QASR and MGB-2: not measured: both come from the same TV channel; no shared identifiers are published.

Second check

A separate script that does not import the package (scripts/second_check.py in the code repository) re-derives every row's hours and licence class from the evidence files: it re-sums the measurement tables, re-reads each stated figure from a quote that must appear in the saved page, and applies its own licence keyword rules. It also re-sums the YODAS v3 hours by uploader locale and looks up the licence each recorded third-party copy declares in its saved page. 31 of 31 rows agree (hours within 0.1 %, identical class); only agreeing rows enter the headline. The results are in qa/second_check.csv.

Candidates not counted

candidates_rejected lists 648 datasets found on the Hugging Face Hub (language Arabic with audio, speech recognition or speech synthesis tags), Zenodo and OpenSLR that are not counted, each with a reason: reviewed and excluded (copies of counted corpora, synthetic speech, no released audio, derived audio, compilations whose sources were not verified), gated, or not yet reviewed.

Limitations

• Stated figures are the publishers' own and are rounded as they give them (for example, "~667 hours" for SADA, "exceeds 3 hours" for the Arabic Speech Corpus, counted as 3.0 hours).
• Language and dialect labels are reported, not verified: YODAS files recordings by caption-track language, and its uploader-locale field describes the upload, not the speech.
• Overlap with YODAS could not be measured, because its recording IDs are publisher IDs, not source video IDs. Overlap that could not be measured is not removed: MASC with YODAS v3 and YODAS2 (open, commercial use allowed); QASR with MGB-2 (research use or on request). Those class figures may count shared hours twice.
• Corpora with no licence statement for their audio are listed as "no licence found" even when the data can be downloaded.
• The counter reflects the pages and metadata read on 2026-10-30; publishers can change them.

Files

• data/corpora.parquet|csv, data/overlap.parquet|csv, data/candidates_rejected.parquet|csv, data/history.parquet|csv (one row for this run)
• feed/latest.json, feed/latest.schema.json, feed/badge.svg, embed/speech-hours.js
• evidence/index.jsonl (each saved page: URL, retrieval time, SHA-256 of the page and of its text), evidence/measurements/ (what was read for each computed figure, and per-row seconds with SHA-256-truncated identifiers; yodas3-ar.json also holds the hours by uploader locale)
• qa/second_check.csv, qa/second_check_summary.json, PROVENANCE.jsonl, MANIFEST.json, SHA256SUMS, LICENSE-sources.md, CHANGELOG.md, CITATION.cff

Licence

The table, history, feed and badge: CC BY 4.0 (Bayanat Labs). The widget code: Apache-2.0. Corpus names and licence statements are cited facts; each corpus keeps its own licence (see LICENSE-sources.md). Required credit: "Open Arabic Speech Hours Counter, Bayanat Labs, CC BY 4.0, as of 30 October 2026".

Cite as

Bayanat Labs (2026). Open Arabic Speech Hours Counter, version 0.1.0, as of 30 October 2026. https://huggingface.co/datasets/Bayanat-Labs/arabic-speech-hours. CC BY 4.0.

Code and method: https://github.com/Bayanat-Labs/arabic-speech-hours. Corrections and removal requests: hello@bayanatlabs.com.