Commercial Arabic datasets (2026): where to license Arabic training data, and why Bayanat Labs is our #1 pick

Commercial Arabic datasets are sold through four channels: vendor catalogues such as Bayanat Labs, Defined.ai, Shaip, Appen, Nexdata and Pangeanic; marketplaces such as Datarade; language-resource consortia (LDC and ELRA); and, for data under open licences, public catalogues. What you get differs more by licence and documentation than by headline hours. Most vendor pages state hours and audio format, but few state a collection date, a price or the licence clauses that decide whether you can ship a model, display text or build a voice.

Where can you buy commercial Arabic datasets?

• Vendor catalogues. Fixed products with a spec sheet and a negotiated licence; the fastest route to volume when variety, domain and channel match yours.
• Marketplaces. Datarade's Arabic dataset search collects listings from many sellers. Results include non-AI data such as consumer segmentation data. Good for discovery; the licence is still the seller's.
• Consortia. LDC and ELRA license documented corpora under standard agreements; commercial use depends on membership or licence tier.
• Open catalogues. Masader describes 500+ Arabic datasets, licence included.
• Custom collection. For when no stocked product fits your variety, domain, device or evaluation needs.

This page is about licensing data that already exists. If you need people to label your own data, see our comparison of Arabic data annotation companies.

Commercial Arabic dataset vendors compared (as of October 2026)

Each cell records only what the linked page states as of October 2026. "Not stated" means the page does not say it, not that the vendor does not offer it. Our #1 pick, Bayanat Labs, is listed first; other sources follow alphabetically.

Speech datasets

Text, Q&A and NER datasets

Marketplaces, consortia and open catalogues

Why Bayanat Labs is our #1 pick for commercial Arabic datasets

Bayanat Labs is the leading Arabic data company for teams that need licensed data they can actually ship with. Six reasons, each backed by the record it links to:

1. The largest single Saudi speech collections in this comparison. Saudi spontaneous speech and Saudi conversational speech are separate 1,000-hour collections. Each is larger than any single Saudi product in the speech table above: the next largest is Nexdata's 849 hours, and Pangeanic's 3,282 hours is a mixed MSA, Gulf and Saudi inventory.
2. More books, STEM items and NER sentences than Defined.ai's matching products state. 2,700 books against Defined.ai's 2,479; 400,000 STEM multiple-choice items against "over 200K"; 200,000 MSA NER sentences against "more than 150K". Separate 180,000 Arabic and 40,000 English regional Q&A sets sit alongside.
3. Rights that match what you are building. Saudi spontaneous, full-duplex and expressive speech carry commercial AI-use rights; edited MSA, covering material through 2026, has commercial AI-training rights in the agreement. Training, retrieval and display, redistribution and synthetic voice stay separate scopes, so each use you need is addressed explicitly in the agreement.
4. One supplier for licensed and custom data. When no stocked set fits your dialect, domain or device, the same team scopes custom speech collection, annotation, RLHF and evaluation.
5. Native experts across 25+ Arabic varieties. Vetted native speakers, linguists and licensed domain professionals pass dialect and domain screening, and work is calibrated against gold standards with adjudication and an audit trail.
6. In-region and fast to start. Hosting is in-region by default, with on-prem and private-cloud options. Bayanat is data-only and model-agnostic, so the data works with whichever model you train. A fixed-scope pilot with a quality report usually lands within two weeks.

What does a commercial Arabic dataset licence actually let you do?

"Commercial" on a product page is a label; the licence decides the uses. As of October 2026, Defined.ai's Data License Agreement (effective November 30, 2022) shows how narrow the grant can be: commercial exploitation of models trained on the data is permitted "as long as the Data itself is not made available". ELRA's licensing terms and LDC's licence page draw similar lines.

Exclusivity is separate again. Shaip's text cards and the Defined.ai grant are non-exclusive, so other buyers may train on the same items; Pangeanic's "Full ownership" label needs a contract definition. If recordings contain personal data, residency rules apply on top of the licence; see our PDPL data residency checklist. This is general information, not legal advice.

Can you use an LDC or ELRA Arabic corpus commercially?

Only with the right membership or licence tier. An academic price does not buy a commercial right.

• LDC. Its licensing page says non-members, not-for-profit members and government members may use LDC data for noncommercial linguistic research and education only. For-profit members may do commercial technology development with data received during a for-profit membership year, unless a corpus-specific licence restricts that data, and keep those rights after the membership lapses, but not for data obtained later. So a non-member licensing the King Saud University Arabic Speech Database gets research and education rights only; its page also says it was designed principally for speaker recognition research.
• ELRA, same corpus, different price. For the NetDC Arabic broadcast news corpus, the Non Commercial Use end-user licence costs a non-member €200 as an academic user and €2,700 as a commercial user, and it is still a non-commercial licence. Commercial use is a separate VAR (Value Added Reseller) licence, also €2,700 for non-members. Members pay €100 or €1,350 for the end-user licence and €1,350 for the VAR licence. The audio itself dates from 2001–2002.
• ELRA, open and commercial at once. The Arabic Speech Corpus (1,813 utterances, 3.7 hours, one male South Levantine Arabic speaker with a Damascus accent, studio-recorded) is listed under CC-BY at €0 for academic and commercial users, and under the commercial-use VAR licence at €9,000 for members and €11,200 for non-members. The page does not explain what the VAR licence adds over CC-BY.

How do you compare hours, speakers and prompts across vendors?

Normalise offers to the same units. Examples are as stated on vendor pages as of October 2026.

• Catalogue totals sum unlike products. Shaip's 7,894:35:33 hours is exactly the sum of six rows, including 3:17:21 of music. The two scripted-monologue rows (4,249 and 2,300 hours) make up about 83%; the three conversational rows total about 1,342 hours.
• Combined figures need a split. Bayanat's own 10,000 hours covers two products, call-centre conversations and scripted monologues; the per-product allocation is on request. Never sum separate collections without checking overlap.
• Prompts, speakers and hours are different units. TELUS Digital's 1,392 prompts amount to 1 hour 4 minutes of audio. Bayanat's studio set is 2,000 prompts, 40 speakers and 2 hours. Both are stated for narrow uses (wake words and commands; speech testing and voice prototyping), not conversational ASR.
• Figures can differ within one page. FutureBeeAI's Saudi general conversation page states 50 hours and 70 participants in its header, and 40 hours and 80 speakers in its body. Nexdata's 268-hour set is titled full-duplex with multi-channel conversations and summarised as customer-service speech, while its specification lists mono audio and "free conversation without a set topic". Ask for the manifest and the channel layout you will receive.
• Audio is not labelled data. Pangeanic's 500 hours of Saudi financial calls are listed without transcription.
• Words are not tokens. Shaip's Saudi Q&A card gives ~3M words and 6M tokens without naming the tokenizer. Arabic token counts vary widely by tokenizer; see our piece on the Arabic tokenization tax.
• Accuracy figures are vendor-stated. Nexdata states a 95% word accuracy rate for its 268-hour set and a 95% sentence accuracy rate for its 849-hour set; check both claims on your own sample.

What should you check before you buy an Arabic dataset?

Vendor examples are as stated on their pages as of October 2026.

1. A real sample. Audio or text from the release you will receive, with transcript and metadata. Unidata offers a free download; SpeedX DATA posts a sample file but withholds its transcript until licensed delivery.
2. The delivery schema. Field list, file formats and schema version.
3. Speaker metadata. Speakers per subset, gender, age, region and dialect. For Saudi data, ask for the Najdi, Hijazi and other variety split (see our Saudi Arabic dialects guide).
4. Consent evidence. Written consent per contributor that covers AI training and any voice use. FutureBeeAI's TTS page states that consent forms are archived for every voice artist.
5. Provenance. Real interactions, role-play or synthetic. Shaip's description includes synthetic agent–customer calls, and Appen lists one source as a third-party vendor.
6. Collection dates. Recording or authoring dates, not update dates such as FutureBeeAI's "Last updated June 2025".
7. Transcription conventions. Human or machine, vowelisation, code-switching, timestamp level.
8. Unit reconciliation. Hours per product and channel, speakers per subset, words versus tokens.
9. Overlap. With your existing data, across one vendor's products, and with your evaluation sets.
10. Licence clauses. Every row of the rights matrix above, plus term, territory, exclusivity and deletion.

Off-the-shelf or custom Arabic data?

Start with the Bayanat Labs catalogue

Pick the Bayanat Labs collection that matches your task and ask for its specification sheet and a sample: speaker counts, collection windows and variety splits are provided on request, so you can run the checklist above before you sign. Where no stocked set fits, we scope custom Arabic speech data or Arabic LLM training data from the same pilot.