AI voice and Arabic dubbing for UAE brands: which Arabic the tools actually publish
A UAE brand has an approved English film and a request, internal or from an agency, to make an Arabic version with a tool. AI Arabic dubbing sounds like one decision with one answer, and the page it gets bought from does nothing to correct that. This article reports what the word Arabic resolves to across five named catalogues, what YouTube already does to an uploaded film without being asked, who has to record a consent statement, what a dub actually is, when a synthetic voice has to be declared, and what this article could not establish. Every statement below was read on the source's own page on 5 September 2026, and each names that page.
AI Arabic dubbing in the UAE starts with a word nobody has pinned down
Which Arabic a Dubai audience responds to is a decision made at brief stage, and the Journal's article on Arabic video production settles that end of it. Its spoken layer ends at human casting. This one begins where that layer is produced by a machine, which is being asked for something the language registries do not treat as a single item.
The registry that owns BCP 47 language subtags classifies Arabic not as a language but as a macrolanguage. It lists thirty individual language subtags underneath it, among them Standard Arabic and Gulf Arabic as separate entries. The ISO 639-3 registration authority classifies both the same way, each an individual living language inside the Arabic macrolanguage.
So a row marked Arabic on a vendor's page is a decision that vendor has made about which of those thirty things it means, and across the five catalogues below that word resolves to between one thing and eighteen. None of those pages is the pricing page.
One note on what follows, carried over from the Journal's article on AI video for a brand film. A tool company's documentation is competent evidence for what that company publishes about its own product. It is not evidence about Arabic, and not evidence about how anything sounds.
The Arabic dub your film may already have
Before anything gets commissioned, there may already be a version.
YouTube publishes that automatic dubbing generates translated audio tracks, that Arabic is among the supported languages, and that videos carrying these tracks are marked auto-dubbed in the description. It publishes that the feature is enabled by default for eligible creators, and that dubs are generated automatically when a new video is uploaded.
What that Arabic track is, in the company's own terms, repays a read. YouTube publishes that seven of the twenty supported languages carry expressive speech, which replicates the original speaker's pitch and intonation, and that Arabic is not among the seven. It publishes that automatic dubs cannot be edited, and that a mistranslation is corrected by fixing the original video's language setting and regenerating. It publishes that dubs may contain errors from mispronunciations, accents, dialects or background noise, and that quality may not be the same across languages.
Two channel settings carry the whole question, one to review dubs before they are published and one to turn automatic dubbing off, which makes this a settings question rather than an edit question. One further detail runs backwards into the shoot: YouTube publishes speech pacing too fast in the original audio as a reason a video is ineligible. The same page publishes lip sync, which alters the speaker's lip movement to match the dubbed audio, as an experimental feature available to select channels.
What the word Arabic resolves to on five vendors' own pages
Watch out
None of that means anybody is lying. It means the buyer opens the vendor's own table, for the product they are buying, on the day they buy it. More Arabic entries than a rival is a fact about a page, not a recommendation.
Five companies, five answers, each read on its own page on 5 September 2026.
Google's is the cleanest instance. The registry that owns region subtags publishes XA inside a range registered as private use, so it is not a country and carries no meaning the standard defines. Google presents that code in a table where every other row names a country, Arabic (XA) sitting above Bengali (India) and English (Australia), without saying what XA denotes. Thirty-eight voices sit there. The number of Arabic places is one.
Microsoft's answer is a tier. All 32 of its prebuilt Arabic voices sit in the Standard tier, and none appears among the 116 at Neural HD, 70 at Neural HD Flash, 15 at Neural HD Omni or 59 Multilingual. Its Style and Roles column is empty for every Arabic voice it lists, which on a brand film is the difference between a read and a performance. Its cloning route publishes 65 locales and the only two Arabic entries there are Egyptian and Saudi. In its video translation table, Emirati Arabic is published as a personal voice source but not as a personal voice target, and it is the only Arabic locale there without a custom lexicon phoneme target, which is the mechanism a brand uses to force the pronunciation of its own name.
Both of Amazon's Gulf voices carry its bilingual marker, while the same company's service card publishes that mixed-language input may produce inconsistent pronunciations or unexpected language switches. ElevenLabs gives two answers, on its text-to-speech page and on the page for the product it sells as dubbing, both open on the same day. No reconciliation is offered here because none is published.
HeyGen goes the other way, which is why none of this is a verdict on AI. Alongside its eighteen Arabic entries it publishes three language counts for the same product on the same day: 175-plus on its developer page, 30-plus on its documentation index, 85 rows in its help centre.
| Company | What it publishes about Arabic |
|---|---|
| One locale, ar-XA, among 61, against four regional variants for English. 38 Arabic voices at that one locale | |
| Microsoft | 32 prebuilt Arabic voices, all Standard tier, none in the four higher tiers, Style and Roles empty on every one. Cloning: Egyptian and Saudi only |
| Amazon | One Arabic voice in the standard engine, a Gulf Arabic pair in the neural engine, neither in generative or long-form |
| ElevenLabs | Text-to-speech page: Saudi Arabia and UAE. Dubbing page: one dialect entry, Egyptian |
| HeyGen | Eighteen Arabic entries on its video translation list, the UAE among them |
Whose voice it is, and who has to sit down and record
Watch out
Every line above is what one company publishes about its own product, binding its own customers. Separate regimes from separate companies are separate contracts, not a rule in this market.
An Arabic version stops being a preference and becomes a diary problem at the point where somebody has to record a consent statement.
Microsoft's is the artefact a Dubai brief runs into. Its voice catalogue and its consent-script file are separate documents that draw the Arabic line in the same place, which is an observation about two documents rather than proof of a policy.
Then the mechanic that decides who books a studio. ElevenLabs publishes that a professional voice clone may only be made of your own voice, that this holds even with the other person's consent, and that the route for somebody else's voice is that they create and verify the clone on their own account and share it. It publishes an estimated three to six hours of fine-tuning after verification, a 24-hour lockout after failed attempts, and that an unverified clone cannot be deleted.
Microsoft, ElevenLabs and HeyGen each put the obligation to obtain the performer's consent on the buyer rather than on itself. What a client owns on an ordinary production is covered in the Journal's article on usage rights and ownership.
Which Arabic is a question with an answer. The tools publish theirs; it is worth reading before anyone records.
OpenAI: a fixed spoken phrase
A mandatory consent phrase published in sixteen languages. There is no Arabic row.
Microsoft: a recorded statement
Published in 49 locales, two Arabic, Egyptian and Saudi, byte-identical, no Emirati entry. Its language must match the training data.
Google: no substitute script
An Arabic consent statement for cloning, filed under the same ar-XA code. No substitute script is permitted.
ElevenLabs: two routes, two strengths
The professional route verifies the voice belongs to the account holder. The instant route takes a self-declared tick at save time.
HeyGen: a video, in any language
A consent video under thirty seconds, identity-matched and validated semantically. Its third level waives collection for accounts holding an indemnity agreement.
What a dub is, from the company that buys them at scale
Before anything is compared to a dub, it helps to know what one is. This rests on a single named commissioner.
Netflix publishes to its production partners that it views dubbing not as merely a language asset but as a production. Casting there belongs to the dub director, is stated to be subjective, and a voice match cannot be prioritised above performance unless a vocal quality is specific to a character trait.
That split decides a brief. The same publisher names the second workflow Simul Sync in its nonfiction guideline, and a corporate or brand film belongs to that family rather than the lip-sync one. Settle which before anyone quotes for either.
One deliverable fact belongs in a schedule. That publisher's Arabic timed-text guide states that where content has been dubbed into Arabic, the dubbing script or dubbed audio is the basis for the Arabic subtitle file for the deaf and hard of hearing. It is derived from the dub rather than translated separately.
Lip-sync dubbing
Adapted to internal lip movements. Every mouth flap accounted for, labials never exchanged for semi-labials unless the mouth is obscured, with the video rather than the original waveform as the main reference.
The voice-over mode
Translated and synchronised to match the original audio timing, without rewriting or adapting to internal lip movements, recorded using standard voice-over practices. A brand film with a presenter or a narrator sits here.
When a synthetic voice has to be declared, and to whom
Watch out
Two publishers landing on the same axis is not agreement and is not a rule. Each sentence above belongs to the company that published it, and to nobody else.
YouTube's position on voice is unusually clean. It requires creators to disclose content generated or meaningfully altered with AI when it appears realistic, then publishes two example lists. Cloning one's own voice to create voice-overs or dubs sits on the list of things creators do not need to disclose. Making it appear as if someone gave advice they did not actually give sits on the list that does require it. Read as what it is, two example lists from one publisher, the axis those examples run along is authorship: whose words they are. YouTube states that neither list is complete, and that qualifier travels with every example.
Meta publishes something of a different shape. It requires people to disclose, using its own AI-disclosure tool, whenever they post organic content with photorealistic video or realistic-sounding audio that was digitally created or altered, and says it may apply penalties for failure. Its label and its advertising rejection are separate and narrower, gated on a particularly high risk of materially deceiving the public on a matter of public importance.
Google's advertising policies sit at the other end. Its Manipulated media and Misrepresentation policies, read on 5 September 2026, name neither audio nor voice and impose no disclosure duty on an advertiser. That is a statement about those two pages on that date.
Underneath the platforms is a layer of contract that may bind a buyer whatever a platform requires. OpenAI's usage policies require a clear disclosure to end users that the voice is AI-generated. Microsoft requires customers to disclose the synthetic nature of text-to-speech voices. AWS states customers should consider disclosing their use of machine learning. Microsoft also publishes the exception a brand will reach for, that where the context makes it clear a synthetic voice is speaking, disclosure may be minimal, momentary or even unnecessary.
What the UAE itself publishes about synthetic voice
Watch out
Awareness guidance is not law. It supports no labelling requirement, no penalty and no procedure.
The federal media regulator publishes under the name National Media Authority, and the former Media Council domain now serves that authority's pages. What effect that change has in law is not established here.
That authority publishes the UAE Media Content Standards in English on its own host: twenty standards, including one requiring an advertisement's content to respect the culture, identity and values of the UAE, and one against material incorrectly attributed to entities or persons. That page, read on 5 September 2026, carries no reference to artificial intelligence, synthetic media, voice cloning, dubbing, labelling or disclosure, and none to the language of content either. That is a claim about that page on that date, not about whether a rule exists elsewhere. The nearest published provision to a rule about synthetic voice is that attribution standard, which turns on whose words a thing is. The operative detail of the federal regime sits in two Cabinet Resolutions published in Arabic only, so an English reader is working from a summary rather than the operative text, and the wider UAE position on AI is covered separately.
The UAE's own AI programme publishes something directly about voice, and it is not a rule. Its Deepfake Guide describes the audible category as voice synthesis and modification, either generating a voice-over creating speech that was never said, or controlling the tone of a person to show unreal emotions. It names the film industry and advertisements among beneficial applications, and states that existing UAE national laws prohibit cyberbullying, actual malice and identity impersonation. On personal data, the UAE government platform publishes that Federal Decree-Law No. 45 of 2021 came into force on 2 January 2022, prohibits processing personal data without the owner's consent apart from stated exceptions, and is available in Arabic only.
What has actually been measured, and what has not
Good to know
Locale coverage is not a quality claim and must not be read as one.
Start with why nothing here is compared with anything else. A peer-reviewed paper at Interspeech 2025 reports that important information, such as the question asked of the listener, is frequently missing from listening-test reports despite affecting how results read. No score below is averaged, ranked or carried across to another study.
There is a published bar. Microsoft's recommended measurements for a usable synthetic voice are a mean opinion score above 4.0, which it states normally requires the human recording to score above 4.5, and intelligibility above 98 per cent. Those measurements attach no language to the figures.
The most useful published measurement explains a failure a brand will meet. The SawtArabi paper at Interspeech 2025 ran a 25-listener panel over Modern Standard Arabic and scored the same text with and without vowelisation, and removing the diacritics lowered all three of its scores. Written Arabic normally omits the short vowel diacritics, so that is the ordinary input case and the mechanism behind a mispronounced brand name. A 2026 preprint from QCRI and HBKU sits against that finding, concluding that more training data compensates for the lack of diacritics to a significant degree, on automatic metrics with no listening test. Neither is reconciled.
On Emirati Arabic, one corpus with a text-to-speech baseline was located, from the University of Sharjah. It was scored by machine transcription rather than by listeners, on two open-source systems, and the system that scored best carries a non-commercial licence, so it cannot be used in a paid campaign.
What we could not establish
Five things this article does not answer, stated plainly.
We have established no price for an AI Arabic dub, a synthetic voice licence or an Arabic voice booking in this market. That is this article's only sentence about cost.
We have established nothing about how any of these voices sounds to an Arabic speaker, because no published listening test of a commercial Arabic voice was located and nothing was tested here.
We have found no published human listening test of Gulf or Emirati synthetic speech in the speech venues, anthologies and preprint archives searched for this article.
We have established nothing about whether a person's voice is personal data or biometric data under the UAE Personal Data Protection Law, because the government platform publishes that law's text in Arabic only.
We have established nothing about how the National Media Authority treats synthetic voice, because its English content standards page and its English services guide address neither.
Common questions
Is our video already being dubbed into Arabic without us asking?
YouTube publishes that automatic dubbing is enabled by default for eligible creators, that dubs are generated on upload, that Arabic is supported, and that these tracks are marked auto-dubbed. Two channel settings govern it: one to review dubs before publishing, one to turn it off.
Which Arabic are we buying when a tool says it supports Arabic?
That depends on the vendor's own table. The registry that owns the language codes treats Arabic as a macrolanguage over thirty individual languages, and the five companies read here resolve the word to between one thing and eighteen, one of them under a region code registered as private use.
Can we use our founder's voice for the Arabic version?
ElevenLabs publishes that a professional clone may only be made of your own voice, even with the other person's consent, and that the route for someone else's is that they create and verify it on their own account and share it. OpenAI publishes no Arabic consent script, Microsoft no Emirati one.
Do we have to tell viewers the voice is AI?
It depends where it runs, and on the contract. YouTube publishes cloning your own voice for voice-overs or dubs on its does-not-need-disclosure list, while stating neither list is complete. Meta requires disclosure of realistic-sounding audio that was digitally created or altered. OpenAI, Microsoft and AWS each publish a term of their own.
Does the UAE have a rule about AI voices in advertising?
Nothing established here answers that. The National Media Authority's English content standards page and its English services guide address neither AI nor synthetic media nor voice on the date they were read, and the operative detail is published in Arabic only. None of that establishes whether a rule exists elsewhere.
Is an AI Arabic voice good enough for a brand film?
No published listening test of a commercial Arabic voice was located and nothing was tested here, so this article does not answer that. What is published is a mechanism. On one 25-listener panel, removing the short vowel diacritics from Modern Standard Arabic lowered the naturalness score.
What to hold onto, in one pass
The word Arabic on a vendor's page carries no fixed meaning. The registry that owns the codes classifies it as a macrolanguage over thirty individual languages, of which Standard Arabic and Gulf Arabic are separate entries, and the vendors resolve it to between one thing and eighteen. One files it under a region code registered as private use. One publishes every Arabic voice in its lowest tier, and Emirati Arabic as a cloning source but not a cloning target. One publishes no Arabic voice in either of its two newest engines. One gives Arabic a single Egyptian dialect tag in the product it sells for dubbing. One publishes eighteen Arabic entries. The consent artefact draws that line again, and it is a recording a named person has to make on a named account. And on the largest video platform the Arabic version may already exist, in the tier without pitch and intonation matching, uneditable, unless somebody changed a setting.
Grolez Films: permissions are scoped before a quote goes out
One position, and nothing beyond it.
Before we quote a job, we work out who has to give permission and who has to apply for it. That reads unusually here, because the consent artefact these suppliers describe is a recording made by a named person on a named account, which is a diary item rather than a line in a licence.
Luis Grolez, the founder, pitches and reviews every job personally. See the work, then get in touch about a specific film and a quarter.
Sources and methodology
Every capability statement here was read on the publisher's own page on 5 September 2026, and each is competent evidence for what that company publishes about its own product and for nothing else. A tool company's documentation is not evidence about Arabic and not evidence about how anything sounds. Five companies publishing five consent regimes is five separately attributed contract terms rather than agreement, and the same holds where two platforms land on the same attribution axis. No score is carried between studies, and the reason is Le Maguer et al. at Interspeech 2025, which reports that the question asked of the listener is frequently missing from listening-test reports. The dub anatomy rests on one named commissioner. The federal regulator's operative texts, its Executive Regulations and its schedule of penalties, are published in Arabic only and were not read, and the UAE's data protection law is likewise published in Arabic only. The body publishing the UAE's media content standards presents itself as the National Media Authority, and the legal effect of that change is not established here. No voice agency, localisation vendor, production-company blog, dubbing studio, law-firm page or news report is cited anywhere, and many were set aside. No price appears, and none was established from the pages read for this article. Every negative here is scoped to the exact page tested on that date. This article is informational and is not legal advice. Prepared 5 September 2026.
- Microsoft, Azure AI Speech language and voice support, for the voice catalogue and its tiers, the Arabic rows, the empty style column, the personal-voice locales and the video translation table: learn.microsoft.com
- Microsoft, personal voice consent how-to and the published verbal-statement locale file, for the consent statement, its 49 locales and the language-must-match rule: learn.microsoft.com and raw.githubusercontent.com
- Microsoft, text to speech transparency note, for the quality bar and the no-language-named observation: learn.microsoft.com, and Microsoft, text to speech disclosure guidelines, for the context-makes-it-clear exception: learn.microsoft.com
- Google, Cloud Text-to-Speech supported voices and languages, for the 61 locales, the single ar-XA Arabic locale and the four English variants: docs.cloud.google.com
- Google, Instant Custom Voice, for the Arabic (XA) row among country-named rows and the Arabic consent statement: docs.cloud.google.com
- YouTube Help, use automatic dubbing, for the supported languages, the default-on behaviour, the expressive-speech list, the uneditable dubs, the settings and the pacing note: support.google.com
- YouTube Help, disclosing use of altered or synthetic content, for both example lists and the not-a-complete-list qualifier: support.google.com
- Google Ads, Manipulated media and Misrepresentation policies, for the scoped negative that neither names audio or voice nor imposes an advertiser disclosure duty: support.google.com
- Amazon, Polly available voices, for the Arabic and Gulf Arabic rows, their engines and the bilingual marker: docs.aws.amazon.com
- Amazon, Polly AI Service Card, for the disclosure recommendation: docs.aws.amazon.com
- ElevenLabs, text to speech and dubbing capability pages, for the two Arabic answers on two pages of one company: elevenlabs.io
- ElevenLabs, professional and instant voice cloning guides, for the own-voice rule, the share route, the timing and the self-declared confirmation: elevenlabs.io
- OpenAI, text-to-speech API guide, for the sixteen-language consent phrase table with no Arabic row and the disclosure requirement: developers.openai.com
- HeyGen, avatar consent page, for the three consent levels and the indemnity waiver: developers.heygen.com; HeyGen help centre, recording your consent video, for the under-thirty-seconds requirement: help.heygen.com; HeyGen help centre, video translation languages, for the eighteen Arabic entries: help.heygen.com
- IANA, Language Subtag Registry, for Arabic as a macrolanguage, the thirty individual subtags and the private-use region range: iana.org
- SIL Global, ISO 639-3, for Standard Arabic and Gulf Arabic as individual living languages inside the macrolanguage: iso639-3.sil.org
- Meta, Community Standards on misinformation, for the AI-disclosure-tool requirement covering realistic-sounding altered audio and the separate label threshold: transparency.meta.com
- Netflix Partner Help Center, Dubbing Guiding Principles and the nonfiction creative guidelines, for the dub-is-a-production statement, the casting rule, the lip-sync criteria and the non-lip-sync workflow: partnerhelp.netflixstudios.com
- Netflix Partner Help Center, Arabic Timed Text Style Guide, for the Arabic subtitle file being derived from the dub: partnerhelp.netflixstudios.com
- National Media Authority, UAE Media Content Standards, for the twenty standards, the attribution standard and the scoped negatives: nma.gov.ae
- u.ae, the official platform of the UAE government, for the data protection law's commencement, its consent requirement and its Arabic-only text: u.ae
- UAE National Programme for Artificial Intelligence, Deepfake Guide, English edition, July 2021, for the audible definition, advertisements as a named beneficial use and the existing-laws sentence: ai.gov.ae
- SawtArabi, Interspeech 2025, for the 25-listener protocol and the vowelisation ablation: isca-archive.org
- Le Maguer et al., Interspeech 2025, for why no score is carried between studies: isca-archive.org
- More Data, Fewer Diacritics, QCRI and HBKU, for the scaling finding reported against the vowelisation result: arxiv.org
- Ramsa, University of Sharjah, for the Emirati corpus and its machine-scored baseline: arxiv.org
- Hugging Face model card, for the non-commercial licence on the system that scored best in that baseline: huggingface.co