ElevenLabs and AI voice over in Arabic: a practical 2026 guide

An AI voice over is narration made by a text-to-speech tool instead of a person in a booth. You paste a script, choose or design a voice, and the tool reads it aloud. ElevenLabs is the name most people search for, and in October 2026 its models read Arabic, as do voices from Microsoft and Google.

For Arabic, the tool is rarely the hard part any more. The hard part is the language: which dialect the brand speaks, how a word without vowel marks will be read, how a price or a brand name sounds, and whether the script was written to be heard at all. Those are a director's decisions, and a listener in Riyadh or Cairo hears the result in the first sentence.

This guide covers what voice tools do, the main options from their official pages, the Arabic details that trip voices up, writing and directing a script, mixing, when to hire a human, and consent, rights and disclosure. It sits inside my guide to turning images into video with AI, where voice is one step of the film.

What AI voice tools actually do

Most voice tools offer four kinds of work, and each carries different risks.

Voice design is where I would point most brands first. ElevenLabs describes it simply: you type a description that can include accent, age, gender, tone and pacing, and it returns three previews. The result belongs to no one, which removes most consent questions before they start.

The main AI voice tools in 2026

These notes come from each company's official pages, checked in October 2026. Models change every few months, so check again before you plan around one feature.

ElevenLabs

ElevenLabs is the voice tool on my software list. On 28 September 2026 it launched Eleven v4 and Eleven v4 Turbo, which its announcement says support more than 90 languages. Its documentation lists Arabic for v4 and for older models it still offers, such as Multilingual v2, which it calls the most stable on long-form generations. Around the models sit Voice Design, Instant and Professional Voice Clones, and Dubbing, whose documentation lists Arabic with a separate Egyptian Arabic option in its newer model.

Microsoft Azure Speech

Microsoft lists its Arabic voices by country: Egypt, Saudi Arabia, the UAE, Kuwait, Qatar, Bahrain, Oman, Jordan, Lebanon, Syria, Iraq, Yemen, Algeria, Morocco, Tunisia and Libya, with a female and a male voice for each. It is built mainly for developers, through an API. A country label tells you the accent a voice was built around, not whether it reads in dialect, so listen before you choose.

Google Cloud Text-to-Speech

Google files its Arabic under the code ar-XA, which its documentation describes as Modern Standard Arabic, with Standard, WaveNet and newer Chirp 3: HD voices. A clear fit for formal narration through an API; a weak fit for a café Reel in Jeddah.

How I would choose

Then check whether the plan includes a commercial licence and whether your credits allow several takes of every line.

Arabic: where AI voices go wrong

Choose the register before you write

Arabic is not one spoken language. Modern Standard Arabic suits corporate, government and educational films, and anything meant for the whole Arab world. Egyptian Arabic speaks to Egypt and carries everyday humour well. Gulf Arabic is not one thing either: a Saudi listener hears the difference between Najdi and Hijazi speech at once. Levantine suits audiences in Lebanon, Syria, Jordan and Palestine.

The register changes the script, not just the accent. "We open at nine" is a different sentence in each. So decide the register first, write in it, and pick a voice whose samples already speak it. The mistake to avoid is the accidental mix: one formal sentence in a street-level ad, or an Egyptian word in a Saudi film.

Add vowel marks where a word can be read two ways

Arabic is usually written without short vowels, so every reader, human or machine, guesses them from context. When the guess is wrong, the meaning changes: the letters علم can be ʿalam (a flag) or ʿilm (knowledge), and ملك can be malik (a king) or malak (an angel).

A case from my own work: the Cairo roastery Abdeen Coffee is بن عابدين in Arabic, and بن here is bunn, coffee. The same letters also read bin, "son of", which turns a coffee brand into a man's name. Written with its vowel marks, بُنّ, the doubt is gone.

You do not need to mark the whole script. Mark the words that could be read two ways, then listen. If the voice still gets it wrong, rewrite the sentence.

Write names the way they sound

Voices stumble most on a Latin brand name inside an Arabic script, or the reverse. Write it in the script of the voice, spelled as it should sound: "SkillUp MENA" becomes سكيل أب مينا. ElevenLabs also supports pronunciation dictionaries with alias rules, which swap a word for a spelling that reads correctly. Keep one per brand, and every new video starts with the names already right.

Numbers, dates and prices

In standard Arabic, numbers agree with the noun in gender and change with grammar, and every dialect says them differently again. ElevenLabs' own best-practice guide recommends writing numbers, dates and currencies out as words before you generate, and in Arabic that advice matters twice. Write prices in words, in your chosen register. Write phone numbers digit by digit. And say which calendar a date belongs to, Hijri or Gregorian, when it matters, as it does around Ramadan and Eid.

Writing a script for the ear

A listener cannot glance back, so every sentence has to land the first time.

• One idea per sentence. Short sentences give the voice places to breathe.
• No brackets, slashes or bullet points. Write them out as words, or leave them out.
• Do not describe the picture. The viewer can see the cup. Tell them what they cannot see.
• Time it aloud. Read at a natural pace with a stopwatch. That is your real length.
• Write each language for its listener. A bilingual brand needs two scripts, not one script and a translation.

Directing the delivery

Generating a voice is closer to directing an actor than to pressing play.

• Cast the voice. Age, warmth and pace should match the brand. Listen to it read your script, not the demo line.
• Shape pauses with punctuation. ElevenLabs notes that Eleven v3 and v4 do not support SSML break tags, so the full stop, the comma and a new line are your pause controls.
• Direct emotion with tags. The newest models accept audio tags in square brackets, such as [whispers] or [excited]. Treat them as direction rather than commands, and test them with your Arabic text.
• Balance steadiness and life. ElevenLabs' Stability setting trades a steady read against a livelier one. A long explainer wants steady; a short ad can take more life.
• Generate in short sections, then choose the best take of each line and join them in the edit.
• Use native ears. Someone from the target country hears what you will miss.

Mixing the voice with music

• Lay the voice first, then cut the pictures and the music to it.
• Choose music with room for words. A sparse instrumental sits under a voice better than a busy track, and lyrics under narration compete for the same ear.
• Lower the music under the voice and let it rise in the gaps. Then check the mix on a phone speaker.
• License the music, and keep the script, voice settings and chosen takes with the project, outside any single tool.

When a human voice actor is the better choice

I would recommend a human voice actor when:

• The voice will be the brand for years, on TV, radio and the phone line. A library voice can be used by anyone; a contracted actor can be exclusive.
• The emotion carries the film: a story of loss, a joke that lives on timing, a character for children.
• The text is religious. Quran recitation and religious texts carry rules and sensitivities of their own, and belong to a qualified human reader.
• The dialect or the poetry is the point. Street-level humour, and metre and rhyme, still need a reader who feels them.

There is a middle path: use an AI voice as a temporary track for timing and client approval, then record a human on the approved script. And if a voice actor wants to offer their voice through AI, ElevenLabs describes a route that keeps it theirs: they create and verify a Professional Voice Clone on their own account, then share it with you privately.

Consent, rights and disclosure

I am a designer, not a lawyer; for a large campaign, ask one. These points I would never skip:

• Never clone a voice without permission. ElevenLabs' Prohibited Use Policy, last updated on 17 August 2026, forbids replicating another person's voice without consent or legal right, or deceiving people about whether a voice was generated by AI. Its Professional Voice Clone can only be made of your own voice, "even with their consent" in its help centre's words.
• Get consent in writing. Before Microsoft builds a professional custom voice, it asks for written permission from the voice talent and a recording of them reading a consent statement. A good standard for any tool.
• Never imitate a public figure, not even as a joke in an ad.
• Read the plan's terms. ElevenLabs says its free plan has no commercial licence and requires attribution when you publish, while paid plans include one. Terms change, so keep a copy with the project.
• Ownership is a separate question, covered in are AI images copyrighted?.
• Disclose where the platform asks. YouTube asks creators to disclose realistic content made or meaningfully changed with AI, such as a real person appearing to say something they did not, and lists cloning your own voice for voice-overs or dubs as needing no disclosure. TikTok asks creators to label AI-generated content that contains realistic images, audio or video.
• Tell the client which voice was used, where it came from, which plan covers it and where it may run.

How voice fits into my video work

The AI video series I direct at SkillUp MENA move through script, stills, motion, voice and edit, with AI as my tool at each step and every creative decision mine. The discipline is the same as for images: decide what you want before you open the tool, direct it clearly, review every output and keep the files in your own folders. I write more about that in AI in graphic design.

Conclusion

An Arabic AI voice over sounds right when the decisions before the tool are right: the register, a script written for the ear, vowel marks where words could go wrong, and names and numbers spelled the way they sound. Then direct the voice like an actor, mix it with care, hire a human when the job needs one, and never use a voice you have no right to.

If you want a brand film with a voice that sounds like your audience, tell me about your project.