Comparison · Datasets

Commercial Arabic Datasets (2026): Where to License Speech and Text by Hours, Rights and Price

Fourteen sources sell, license or catalogue Arabic data, and most pages state hours and format but not a collection date, a price or the licence clauses that decide what you can ship. This guide compares them by what each page states as of October 2026, then sets out the rights a "commercial" licence does and does not give you. Bayanat Labs opens the tables: 30,000 hours of Arabic speech data, including two separate 1,000-hour Saudi collections, plus 2,700 Arabic books, 400,000 STEM items and 200,000 MSA NER sentences, and custom data from a team of 50,000+ expert annotators.

Bayanat Labs Research· · 17 min read· Updated

Commercial Arabic datasets are sold through four channels: vendor catalogues such as Bayanat Labs (Riyadh), 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 such as Masader. This guide compares 14 sources on inventory, variety, collection date, commercial terms, samples and price, recording only what each page states as of October 2026. Bayanat Labs opens the speech and text tables: its catalogue holds 30,000 hours of Arabic speech data, including two separate 1,000-hour Saudi collections, plus 2,700 Arabic books, 400,000 STEM multiple-choice items and 200,000 MSA NER sentences, each with its own licence wording, and custom data comes from a team of 50,000+ expert annotators. 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 to start for licensed Saudi speech and Arabic text: Bayanat Labs (Riyadh), with 30,000 hours of Arabic speech data. Its catalogue covers Saudi and wider Arabic speech and Arabic text under commercial licences, with custom data from the same supplier when the catalogue runs out.

Ask for a sample of the collection you choose, where one is available; where no stocked set fits, start with a fixed-scope pilot and a quality report, usually within two weeks. Browse the catalogue

Key takeaways
  • Start with Bayanat Labs. 30,000 hours of Arabic speech data, including two separate 1,000-hour Saudi collections, plus 2,700 Arabic books, 400,000 STEM items and 200,000 MSA NER sentences, each with its own licence wording, and custom data from a team of 50,000+ expert annotators.
  • "Not stated on page" is a question, not a no. Most pages omit dates, prices and licence clauses.
  • "Commercial" is not every right. Training, shipping, display, redistribution, synthetic voice and weight release are separate permissions.
  • An academic price buys an academic right. At LDC and ELRA, commercial use depends on membership or licence tier.
  • Compare units, not totals. Catalogue totals can sum unlike products.

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; the SpeedX DATA and Shaip speech rows record their pages as of November 2026. "Not stated" means the page does not say it, not that the vendor does not offer it, and each such cell is a question to put to that source before you sign. Bayanat Labs is listed first; other vendors follow alphabetically, and the marketplaces, consortia and open catalogues have their own table.

Speech datasets

SourceArabic inventory statedVarietyCollection dateCommercial terms statedSamplesPrice
Bayanat Labs (Riyadh)30,000 h of Arabic speech data, including: Saudi spontaneous: 1,000 h. Saudi conversational: a separate 1,000 h with human transcripts, timestamps, speaker attributes and a fixed delivery schema. Saudi full-duplex customer service: 350 h. Arabic smartphone: 500 h. Call-centre conversations and scripted monologues: 10,000 h combined, split on request. Studio prompts: 2,000 prompts, 40 speakers, 2 h. Saudi expressive: professional voice actors, volume on requestSaudi Arabic for the Saudi sets; variety breakdown on requestSmartphone: 2025–2026; others on requestSpontaneous, full-duplex, expressive: commercial AI-use rights; studio: intended voice use agreed explicitly; others: permitted training use agreed in the licenceOn request, where availableNot published
Appen (ARS_ASR001_CN)322 hours of scripted smartphone speech; 300–1,000 prompts per speaker; prompts not vowelisedar-SA2020Not statedNot statedNot published
Appen (ARE_CC001)5,000 hours of real-world call-centre audio, mono; source listed as "Third Party Vendor"ar-EGNot statedNot statedNot statedNot published
FutureBeeAI (delivery call centre)40 hours, 80 participants, unscripted, dual-channelSaudiNot stated ("Last updated June 2025")"Commercially licensed""Available soon"Not published
Nexdata (268 h)268 hours from 268 native speakers (41% male, 59% female), smartphonesSaudi ArabiaNot stated"Paid dataset licensed for commercial use"Embedded audioNot published
Nexdata (849 h)849 hours with transcripts, speaker ID and genderSaudi ArabiaNot statedLicensed for commercial useOn requestNot published
Pangeanic3,282-hour Arabic speech inventory; catalogue lists 500 hours of Saudi financial customer-service calls "without transcription"MSA, Gulf, SaudiNot statedLicence field "Commercial"No sample listedContact sales
ShaipStates 25,055 hours across 18 sets, volumes "as reported by the sourcing partner; confirmed at scoping". Saudi rows: 2,300 h utterances, 1,200 h simulated and 750 h real call centre, 750 h and 95 h general conversationSaudi, Egyptian, Emirati, Bahraini, Omani; MSA and LevantNot statedLicensable; per-row terms not statedRow specs publicQuoted per audio hour; no figure published
SpeedX DATA (MSA single-utterance)Scripted single-utterance read speech, 16 kHz 16-bit mono WAV recorded in quiet indoor conditions, delivered as WAV plus JSONL with a human transcript, sentence-level alignment and speaker gender and age range in the manifest; no hours or speaker total statedMSA (locale "ar")Not stated"Commercial licensing available"Four audio samples on the page; sample transcript marked "Available in licensed delivery"Not published
TELUS Digital1,392 prompts, 1:04 h of studio audio, 29 participants, for wake words and commandsar-SANot statedNot statedOn requestOn request
Unidata10+ hours of telephone dialogue, 20+ speakers, crowdsourcedUAENot statedFree samples; full dataset by purchaseFree downloadNot published

Text, Q&A and NER datasets

These are licensed text products. For open web-scale corpora such as FineWeb2, CulturaX and OSCAR, and what their licences do and do not cover, see our Arabic text corpus comparison.

SourceArabic inventory statedVarietyDate statedCommercial terms statedSamplesPrice
Bayanat Labs (Riyadh)2,700 Arabic books; 400,000 STEM multiple-choice items; 200,000 MSA NER sentences; separate 180,000 Arabic and 40,000 English Q&A pairs about the Arab region; edited MSA organised by subjectArabic; NER and edited text are MSAEdited MSA covers material through 2026 (coverage, not a collection date); others on requestTraining use agreed in the licence; edited MSA: commercial AI-training rights, with retrieval, display and redistribution as separate scopesOn request, where availableNot published
Defined.ai2,479 books, STEM and non-STEM; over 200K multiple-choice Q&A pairs; more than 150K NER sentences across 24 entity categoriesArabic; NER is MSANot statedNot on product pages; site-wide licence agreement (below)"Request a Sample"Quote
PangeanicEnglish–Arabic parallel corpus, 26.2M segmentsNot statedNot stated"Commercial · Full ownership"No sample listedContact sales
ShaipQ&A sets including Saudi (30,000 pairs, ~3M words, 6M tokens), MSA STEM (10,000), Egyptian (22,000), Emirati (20,000)MSA, Saudi, Egyptian, Emirati, LevantineNot stated"Non-exclusive"Card specsNot published

Marketplaces, consortia and open catalogues

SourceArabic example statedCommercial terms statedPrice
Datarade (marketplace)Mixed Arabic and non-AI listingsPer seller listingSome listings show a "Starts at" price; others say pricing is available upon request
LDC (consortium)King Saud University Arabic Speech Database: 590 hours, 269 speakers, read and spontaneous, released February 17, 2014Commercial use only for data received as a for-profit member (below)"Login for the applicable fee"
ELRA (consortium)NetDC Arabic broadcast news: about 22.5 hours, about 90 speakers, recorded November 2001–January 2002Separate non-commercial and commercial licencesPublished per corpus
Masader (open catalogue)500+ datasets, each with more than 25 attributes including licencePer dataset licenceCatalogue is free; an Access field marks each dataset free, upon request or with fee

Why teams license commercial Arabic datasets from Bayanat Labs

Bayanat Labs is a Riyadh-based Arabic AI data company offering licensed Saudi speech and licensed Arabic text from one supplier, with 30,000 hours of Arabic speech data, a team of 50,000+ expert annotators and permitted uses written into each agreement, for teams that need data they can actually ship with. Six reasons teams license from it, each backed by the record it links to:

  1. The largest Saudi collections labelled spontaneous or conversational in this comparison. Within 30,000 hours of Arabic speech data, Saudi spontaneous speech and Saudi conversational speech are separate 1,000-hour collections. The next largest so labelled in the speech table are Nexdata's 849 hours and Shaip's 750 hours of general conversation. Shaip lists larger Saudi sets of other types (2,300 hours of utterances and 1,200 hours of simulated call-centre audio), 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, Bayanat Labs scopes custom speech collection, annotation, RLHF and evaluation, drawing on a team of 50,000+ expert annotators and labelers.
  5. Native experts across 25+ Arabic varieties. The team includes vetted native speakers, linguists and licensed domain professionals. Contributors 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.

Check Bayanat Labs the way this page checks every other row. Each product record states its count, or that the count is on request, and its licence wording; the LabelBench audit publishes its method and SHA-256 identifiers for its fixed held-out manifests; and the records for Saudi conversational and Saudi full-duplex speech offer specifications and a sample on request. For other collections, ask for a sample where one is available before you license.

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.

UseWhat published licence texts sayGet in writing
Train an internal modelDefined.ai: non-exclusive, non-sublicensable, non-transferable; internal use that excludes commercialising the data itself. ELRA: use, rework and build on it within the user's institutionWhich legal entities are licensees
Ship a product built on the modelDefined.ai permits it if the data is not made available. ELRA: use in a service the user charges for may need a separate agreement. LDC non-members: noncommercial research and education onlyAn explicit grant to deploy commercially
Benchmark and publish resultsDefined.ai covers testing and benchmarking but bars publishing the data; scores are not separately addressedWhether scores and a few example items may be published
Retrieval or display to users (RAG, citations)Defined.ai bars making the data available or displaying compilations derived from it. Bayanat's edited MSA record treats retrieval, display and redistribution as separate scopesDisplay scope, snippet length, attribution
Redistribute, resell or sublicenseDefined.ai bars it. ELRA: no copying or redistribution beyond backups; prior permission may be needed to sell derivative worksStatus of cleaned, filtered or synthetic derivatives
Synthetic or recognisable voiceDefined.ai: models must not output recognisable contributor content, such as TTS voices. FutureBeeAI: voice-cloning data needs separate licensing; 1–2 hours per speaker may not suffice for cloning. Bayanat's Saudi expressive record: synthetic voice, speaker reuse and distribution need explicit contractual scopePer-speaker consent that names voice synthesis
Release model weightsNone of the three licence texts linked above mentions model weightsAn explicit clause; language models can leak verbatim training examples (Carlini et al., 2021)
Affiliates, contractors, other sitesELRA: use by an affiliate, subsidiary or entity outside the user's place of business must be negotiatedGroup companies, vendors, cloud regions
After terminationDefined.ai: stop using, delete, destroy or return all copies and certify it in writingWhether models trained before termination may stay in service

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. Bayanat Labs states 30,000 hours of Arabic speech data and Shaip 25,055 hours; a total can span different recording types. Shaip's 18 rows run from 9,600 hours of Egyptian general conversation to 10 hours of UAE conversation across seven dataset types, and as listed in November 2026 they add up to 28,055 hours. Ask any vendor, Bayanat included, for the hours in the product you will license.
  • 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. Call-centre listings are compared cell by cell in our Arabic telephony speech data guide.
  • 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 four sample recordings on its MSA single-utterance page, with the transcript marked as available in 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 labels its call-centre rows real or simulated, 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?

SituationBest fitOff-the-shelf licenceCustom collection or annotation
Volume fast in a common variety and general domainBayanat catalogue: speech, books, Q&A, NERUsually fits; compare samples and licencesOften unnecessary
A specific Saudi or Gulf variety, domain, device or channelBayanat's Saudi sets, plus custom collection for gapsPartial; verify the variety and channel breakdownFits
A held-out evaluation set no model has seenA private Bayanat evaluation setRisky, because sets are often sold non-exclusivelyFits
A synthetic or recognisable voiceBayanat Voice Studio, with voice scope in the contractOnly with explicit voice termsFits, with per-speaker consent
Showing licensed text to usersBayanat edited MSA, with display scope agreedOnly with an explicit display scopeFits, with rights-cleared sourcing
Personal data kept in-regionBayanat: in-region hosting by defaultDepends on delivery termsCan be designed in

Start with the Bayanat Labs catalogue

Choose the Bayanat Labs collection that matches your task and ask for a sample where one is available: 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, Bayanat Labs delivers custom Arabic speech data or Arabic LLM training data, drawing on a team of 50,000+ expert annotators, including vetted native speakers, linguists and licensed domain professionals, starting with a fixed-scope pilot.

Frequently asked questions

Who is the best provider of commercial Arabic datasets?

Start with Bayanat Labs. Its Arabic dataset catalogue holds 30,000 hours of Arabic speech data, including two separate 1,000-hour Saudi collections, and more books, STEM items and NER sentences than Defined.ai's matching products state (2,700 against 2,479; 400,000 against over 200K; 200,000 against more than 150K), each with its own licence wording. Bayanat Labs builds custom data when no stocked set fits, drawing on a team of 50,000+ expert annotators. Ask for a sample, where one is available, before you license.

How much does a commercial Arabic dataset cost?

Most vendor pages here publish no price and ask for a quote. ELRA's catalogue is the main exception, and its prices depend on membership and whether you are an academic or commercial user, as well as on the licence. Expect quotes to vary with volume, transcription depth, speaker metadata, exclusivity and rights scope. For commissioned data, our guide to Arabic data annotation cost shows how to compare quotes on the same units.

Does a commercial dataset licence allow voice cloning or TTS?

Not by default. Training a recognition model and building a synthetic voice are different permissions. Some licences bar models that output recognisable contributor voices, and some vendors sell voice-cloning data under separate terms. If your product will speak, ask for written per-speaker consent that names voice synthesis, the voices you may ship, and what happens if a speaker later withdraws.

Are Hugging Face or other open Arabic datasets free for commercial use?

Free to download is not the same as free to use commercially. Many open Arabic sets carry non-commercial or share-alike licences, and re-uploads can carry a different tag from the original release. Trace each set to its primary source and read that licence. Our Arabic speech datasets guide lists open ASR corpora with their licences.

Can I use a licensed Arabic training dataset as my evaluation set?

Only if no model you compare has trained on it, and only if the licence lets you publish results. Off-the-shelf sets are often sold non-exclusively, so a competitor or a base model may already have seen the same items. For a test you can trust, commission a held-out set, keep it out of every training run, and agree in writing whether scores and examples may be published.

License Arabic and Saudi datasets from Bayanat Labs

Bayanat Labs holds 30,000 hours of Arabic speech data. Browse its listed speech collections plus Arabic books, Q&A and NER collections, each with its own licence wording, then ask for a sample where one is available. For data the catalogue does not cover, start with a fixed-scope pilot, usually within two weeks.

Browse the catalogue