How to make voiceovers in Arabic, Hindi and English with one tool

This page may contain affiliate links from which Blur Studio earns commissions.

A buyer who doesn't hear their own language in your marketing is a buyer who may not engage with it at all, regardless of how good the content is. Multilingual voiceover solves this, but only if you pick a model that actually covers every language you need, which isn't automatic, since ElevenLabs runs several different speech models with different language lists.

Which model actually covers which language

This is the detail that trips people up: ElevenLabs' models don't all support the same languages. Eleven Multilingual v2, the model most commonly used for narration, covers 29 languages including Arabic, but Hindi is not one of the languages it names directly. If your project needs Arabic, Hindi and English together from one consistent workflow, Eleven v4, which expanded to more than 90 languages when it launched in September 2026, is the model that actually covers all three. Before you commit to a workflow, check the specific language you need against the specific model's current language list on ElevenLabs' own documentation, since this detail changes as models update and isn't something you can assume carries over from one model to the next.

How to do this properly

1. Get your script translated properly first, not machine-translated and used as-is. A voiceover reads out whatever text you give it exactly as written, so any translation error or awkward phrasing in the source text gets voiced too.
2. Have a native speaker of each target language review the translated script before you generate any audio, catching phrasing issues while they're still easy to fix in text, rather than after you've spent credits generating audio from a flawed script.
3. Generate each language's version using ElevenLabs with the model that actually covers that language (check this per language, not once for the whole project).
4. Listen specifically for names, numbers and addresses, which is where multilingual voiceover most often slips, a street name, a price, a property reference number read with the wrong emphasis or accent.
5. Have the same native speaker who reviewed the script also review the final audio, since reading a translation and hearing it spoken are different checks, and catching an error at the audio stage is still cheaper than publishing it.

The specific Arabic and Hindi pronunciation issues worth knowing

Arabic script normally omits short vowel marks (tashkeel/diacritics) in everyday writing, which humans fill in from context but which can genuinely change how an AI model reads a word, since the same unmarked word can be read multiple correct ways depending on the missing vowels. For anything where precision matters (a legal term, a brand name, a less common place name), adding the diacritics back into the script before generating the voiceover removes that ambiguity, rather than leaving the model to guess the most likely reading. For Hindi, the more common issue is numerals and dates: Hindi text frequently mixes Devanagari script with Western Arabic numerals (writing "25" rather than "२५"), and how a model reads those numerals, as a Hindi number or an English one, can sound subtly foreign if it defaults to the wrong convention. Listen specifically to how numbers, dates and prices are read in your Hindi audio, not just whether the words around them sound natural.

A related, very real issue for Dubai-focused content specifically: code-switching, where an English brand name, building name or neighborhood ("Downtown Dubai," "Business Bay") sits in the middle of an otherwise Arabic or Hindi sentence. Models handle this with varying success, sometimes reading the English phrase with a foreign-sounding accent inside an otherwise natural Arabic sentence, which is jarring to a native listener even when every other word is correct. If a script has several of these embedded English terms, test that exact sentence structure specifically rather than assuming it'll sound as natural as a sentence entirely in one language. Where a specific name or term consistently reads wrong, a pronunciation dictionary (a feature ElevenLabs offers separately for locking in exactly how a specific word should sound) is the fix, worth setting up once for any name or term you'll use repeatedly rather than re-checking it every time. The steps are in fixing Arabic place and developer names in listing audio.

A worked example: one listing, three languages

Take a single property listing that needs to go out in Arabic, Hindi and English for a Dubai-based audience. The process in practice: write the English script first, since it's usually the source everyone reviews and signs off on. Have it professionally translated into Arabic and Hindi, not machine-translated, with the translator specifically flagging any embedded English names (the building name, the area, any amenity brand names) that should stay in English within the translated sentence rather than being translated themselves, since translating a proper noun usually sounds wrong. Have a native Arabic speaker and a native Hindi speaker each review their translated script before any audio gets generated, catching awkward phrasing and confirming which embedded English terms should actually stay in English. Generate all three using Eleven v4 for consistency, since it's the one model that covers all three languages properly, rather than switching between Multilingual v2 for Arabic and v4 for Hindi, which would mean managing two different workflows for one listing. Finally, have the same two native speakers listen to the finished audio specifically for the numbers (the price, the square footage, the floor number) and the embedded English names, since those are the two places this kind of content most often goes wrong, even when the rest of the sentence sounds completely natural. If the page carrying the audio also needs both languages, building a website in two languages covers that side.

What this costs

Pricing doesn't change by language, ElevenLabs charges the same per-character rate regardless of which supported language you're generating, $0.08 per 1,000 characters on standard models like Multilingual v2 and v4, or half that on faster models where available. A typical property listing script of 150 to 250 words runs to roughly 1,000 to 1,500 characters per language, so producing the same script in three languages costs about three times a single-language generation, not an unusual multiplier, just worth knowing if you're budgeting for several languages at volume.

My honest take for Dubai-focused marketing specifically

If you're marketing property or services to a genuinely multilingual buyer base, the ROI on getting this right is higher than most businesses assume, a buyer who hears their own language delivered naturally is meaningfully more likely to engage than one reading subtitles or hearing an obviously foreign accent reading their language. Where I'd push back: don't assume "the AI covers the language" means "the AI sounds natural in the language" without actually checking. Model quality varies noticeably across languages, some languages get more training data and sound more natural than others, and the only way to know is to listen to a real sample in each specific language you need, not just confirm it's technically supported. For video, see when AI dubbing is enough and when to hire a native speaker.

The other honest point worth making: regional accent matters more than most people assume when picking a voice for Arabic specifically. Gulf Arabic, Egyptian Arabic and Levantine Arabic sound noticeably different to a native speaker, even though they're technically the same language, and a voice trained with a heavier accent toward one region can read as subtly off to a Dubai-based audience even when every word is pronounced correctly. If the voice you're using has a selectable accent or dialect option, pick deliberately rather than defaulting to whichever Arabic voice appears first in the library (see finding a native-sounding Arabic or Indian English voice), and if it doesn't, have a Gulf-based native speaker specifically confirm the accent reads naturally for your audience, not just that the Arabic itself is correct.

When to build something more structured

For a handful of listings or campaigns, generating each language's audio by hand through the dashboard is reasonable. If you're running this across an entire property portfolio or a regular stream of new listings, each needing the same script structure translated and voiced into multiple languages, that's worth automating: a template that takes one source script and produces correctly-voiced audio in every language you support, through the API, without someone manually running each language through the dashboard separately. That's the kind of multilingual content pipeline I build for clients with real volume here.

What that actually looks like: one script enters the system, a translation step (either a connected translation service or a human translator's output pasted in) produces the Arabic and Hindi versions, each gets voiced automatically through the right model with the right saved pronunciation dictionary already applied for recurring names, and the three finished audio files come out the other end without anyone manually running the dashboard three times per listing. The review step, a native speaker checking each language, still matters and shouldn't be automated away, but everything mechanical around it can be, which is usually where the actual time cost was hiding once you're producing more than a handful of listings a week.