Julie Fry Headshot Hi Res
Julie Fry 8 min read

ElevenLabs v4 in Arabic, Spanish, Greek and Mandarin: a side-by-side test

Generating an AI voice-over is easy. You just choose a voice, paste in your script and press a button. A few seconds later, you have an advert in a polished-sounding voice enthusiastically selling a romantic weekend in Buckinghamshire. Anyone who has used AI voice-over on real projects knows that first take is only the start. The real question is how far that first take is from something a native speaker would accept, and how much that gap varies from one language to another.

A woman wearing headphones relaxes on a sofa while listening to audio, holding a blue mug and smiling as she looks into the distance.

We started with a short radio advert, translated it into Arabic, Latin American Spanish, Greek and Mandarin Chinese, then used ElevenLabs v4 to voice each translation.

The test focused on three areas where AI voice-over typically runs into trouble: how natural it sounds to a native speaker, how it handles prices, dates, abbreviations and English names inside a translated script and whether it can perform a script rather than simply convert it to speech.

The test: one advert, four languages

We chose a fictional hotel radio advert because it let us squeeze quite a lot into around 75 seconds. You can read the sample script here.

There were facts to communicate, but there was also something to sell. The opening invited a couple to imagine escaping to a hotel in the countryside. Then came the practical stuff: room type, meals, facilities, price and booking deadline. Finally, the script tried to paint a romantic picture of candlelit dinners, moonlit walks and breakfast in bed. That meant the voice had to change gear several times without sounding as though three different people had wandered into the recording booth.

We also included elements that are typically tricky for AI voices, such as place names, brand names, abbreviations, prices and percentages: Buckinghamshire, River & Rose Rewards, VIP, Wi-Fi, £299 and 20%. The aim wasn't to trip the AI up, though. The translations were prepared in the usual way. If our translators felt something would naturally be translated or transliterated, it was. English terms remained only where it made sense to keep them.

First, choose your voice

ElevenLabs' Voice Library contains voices shared by its community, including verified Professional Voice Clones created from real speakers' voices. The library can be filtered by language, accent and use case. To keep the trial fair, we selected a male voice that was listed under marketing and advertising for each language to allow for easier comparison.

The voice options on the platform aren’t equally broad in every language. Latin American Spanish gives you plenty of clearly described options. Greek offers fewer voices, and many of the Chinese voices aren't divided into categories, which can make selection a little trickier.

There is also an element of judgement that no filter can help with: what should a Chinese hotel advert actually sound like? Would listeners respond better to something lively and animated, or calmer and more polished? If a Mandarin voice sounds rather neutral to an English ear, is it a dull choice or actually a very suitable one for what you are selling? 

Directing the performance with tags

ElevenLabs lets you put instructions directly into the script to tailor the performance.

Some are simple:

  • [softly] 
  • [cheerfully] 
  • [thoughtful] 
  • [short pause]

Others can be more descriptive with multiple cues:

  • [warm, inviting] 
  • [brighter, promotional] 
  • [slower, intimate]

They are snippets of direction for the AI VO. You just drop them in at the beginning of the appropriate sentence.

ElevenLabs also has an Auto Tag function. Give it the untagged script and the system decides for itself where those instructions should go. The different options gave us the basis for the three test recordings in each language.

  • A: No tags
    We let ElevenLabs interpret the script for itself.
  • B: Auto Tag
    ElevenLabs decided where and how the performance should change.
  • C: Human-directed tags
    We decided where the tone, pace and emotion should change, created our own tags and dropped them in.

Our working assumption was that quality would improve from versions A to B to C, with human-directed tags producing the best result. But that is not quite how things unfolded.

Same script, same voice, different take

In ElevenLabs, you can regenerate the VO as many times as you like (assuming you have the credit). However, pressing Generate again for the same script and voice using the same settings won't necessarily give a better performance, just a different one. And that distinction is important. Generate the VO a second time and the opening might suddenly sound better while the ending gets worse, or the pronunciation of a word might improve while the pacing changes somewhere else. So, it’s worth keeping copies of all iterations in case you want to roll back.

For the purposes of our test and in the interest of fairness, we generated each version of the script just once rather than repeatedly hitting Generate until we happened upon a favourite.

Our native reviewers, all with VO experience, then listened to the three recordings blind. They knew they were assessing AI-generated audio, but they were not told which platform had been used or how versions A, B and C had been produced and what was different about them.

They scored pronunciation, naturalness, emotional delivery and pacing, giving each recording a total out of 20.

You can listen to all 12 recordings and view the scripts here to judge for yourself
Arabic: A, B, C
Mandarin Chinese: A, B, C
Greek: A, B, C
Latin American Spanish: A, B, C

The results

Results table. Arabic: A 19/20, B 18/20, C 17/20. Latin American Spanish: A 14/20, B 15/20, C 17/20. Greek: A 9/20, B 10/20, C 14/20. Mandarin Chinese: A 9/20, B 14/20, C 16/20.

Three languages broadly behaved as expected. As you can see, Arabic politely tore up the hypothesis.

Arabic: when helping made things worse

That 19/20 needs a little context. Our reviewer wasn't saying the untagged Arabic was flawless or ready to publish untouched. He scored pronunciation 4/5 because he found several specific errors and classed it as suitable for client delivery with minor changes. But he gave it full marks for naturalness, emotional delivery and pacing.

The performance itself was remarkably good. Then the tweaking started: Auto Tag made it slightly worse, and our own human direction made it worse again. In B, the voice suddenly dropped into an almost whispered delivery, while C introduced another abrupt tonal change.

After we revealed how the three versions had been produced, our reviewer went back to his notes and said:

“Left to decide the emphasis itself, it seems to have produced a more balanced, natural-sounding performance.”

It's tempting to assume AI can be improved with more prompting. Here, the original performance was already making sensible choices and our attempts to improve it simply pushed those choices too far.

The pronunciation errors had a different cause. They stubbornly survived through versions A, B and C, so the emotion and delivery tags weren't the culprit. The problem was in the Arabic script itself. Ordinary Arabic writing generally leaves out many of the short-vowel and pronunciation marks that tell you exactly how a word should be read. A human understands the grammar and meaning and supplies them instinctively. The AI had to make a choice, and in several places it chose the wrong option.

Our reviewer's recommendation was clear:

“The best results come from a native linguist preparing a fully vocalised script and then reviewing the AI output before delivery.”

Put simply, that means adding the missing Arabic vowel and pronunciation marks before asking the AI to read the script. For Arabic, the most valuable human intervention may happen before generation rather than through increasingly elaborate instructions about how the voice should sound.

Spanish and Mandarin: right mood, but too much of it

Spanish and Mandarin both improved once direction was introduced, but Auto Tag revealed an interesting habit: it often recognised correctly where the mood ought to change without having quite the same judgement about how far to take it.

Our Spanish reviewer described one transition as:

“An under-projected whisper, followed by an abrupt, jarring leap into a loud, high-energy corporate tone.”

The AI had grasped the structure of the advert, with an intimate section followed by a promotional one. It had just turned the knobs rather too far and jumped from one tone straight to the other.

Our Mandarin reviewer heard much the same thing:

“Suddenly turns breathy/whispered — a strange, inappropriate and over-dramatic shift.”

Elsewhere, the AI added a laugh that nobody had asked for, which was actually a little creepy.

The human-directed C versions were better in both languages because the emotional changes were smoother.

If we brief a professional voice artist to record a romantic hotel advert, we don't have to annotate every sentence with warm here, more promotional now, pause, slow down, be intimate again. VO artists are actors. They read the script, understand what it is trying to achieve and make those decisions themselves. We may give an overall brief, or ask for a line to be warmer, softer or less salesy, but much of that interpretation is exactly what we are paying them for. With AI, some of the acting-skill requirements shift upstream to the person preparing the input.

Greek: when perfect English was too perfect

Greek produced a different problem. Every time an English name appeared, the AI seemed to jump briefly into an extremely English pronunciation before returning to Greek. The English itself was excellent, and that was the problem.

Our Greek reviewer explained that, as you would imagine, Greek speakers naturally retain something of a Greek accent when using English words and names. The AI's sudden native-English delivery therefore sounded conspicuous inside the surrounding Greek. She described it as “too perfect”.

Her overall verdict was:

“The AI knows how to speak Greek. What it still needs is to understand what we're trying to make the listener feel.”

Elsewhere, she described the Greek itself as technically impressive but still “missing the soul” of how a Greek speaker would naturally perform it. A voice can pronounce every word correctly and still not sound quite right, and that was one of the clearest findings of the test.

Mandarin: finding the fault was the easy part

Mandarin gave us one small but remarkably persistent fault: an unnatural break in “早、晚餐”. It appeared in A, survived Auto Tag and was still there in C, and our reviewer called it the most obvious AI giveaway in every version.

Identifying the problem was easy, but fixing it was another matter. More tags clearly weren’t going to help. The options were to change the punctuation, rewrite that part of the Chinese, try a pronunciation tool, or hit Generate again and hope. As previously mentioned, another generation is a whole new performance, which might solve this problem but introduce a new one somewhere else.

ElevenLabs does offer more targeted regeneration in Studio, where a selected phrase or sentence can be generated again rather than recreating the entire recording. But someone still needs to know what to change and how to change it.

Our Mandarin reviewer has worked on a number of AI-generated Mandarin projects and described the most reliable workflow as:

“Have a human review the generated audio, identify and adjust the problematic sections in the script, and then regenerate those parts.”

In practice, the Mandarin linguist or VO reviewer identifies the problem and determines what needs to change in the Chinese script, rather than a non-Chinese-speaking project manager trying to repair it in ElevenLabs. The revised wording or punctuation can then be put back into the system and that section regenerated. So, the reviewer's role extends beyond checking the finished product to helping shape the next version.

It's also worth noting that a native speaker isn't automatically a VO reviewer. Think of it like reviewing a film. Most people can tell you what they didn't like, whether the plot dragged or the ending fell flat, but far fewer could tell you how to fix it. A native speaker can hear that a VO sounds off. But if the problem is intonation, acting, emphasis, pacing or whether the voice suits an advert, you need somebody with the ear and experience to diagnose the performance and say what would actually make it better. Somewhat ironically, improving an AI voice-over may sometimes require the skills of a human VO artist.

Time and cost

There is no question that AI wins hands down on the speed of producing the first recording, which takes a few seconds. But that recording is rarely the finished product, and our test files were only around 75 seconds long.

We've recently been involved in a project using AI VO for a webinar of around 90 minutes. A native reviewer listening to an hour and a half of audio may return comments scattered throughout the recording, and someone then has to work through them one by one. Each might need a script change, a pronunciation fix or different direction, and any regenerated section may need to go back to the reviewer again. Add a project manager, native reviewer and client stakeholders into the mix and the workflow expands and must be repeated for each new batch of takes. The actual generation may take seconds, but the production cycle can stretch across days, weeks or longer.

A short script recorded by an experienced human VO artist may, in contrast, come back as a strong performance after the first take, followed by a small number of pickups if anything needs changing.

AI may still be faster overall, but the answer depends on which bit of the process you are timing and how much material there is to fix. AI removes much of the traditional talent cost and potentially large usage or buyout fees, but some of those savings can reappear elsewhere in hourly costs for project management and native-speaker reviews. It may still come out considerably cheaper, but the fact that the audio cost a few pounds to generate doesn't mean the finished VO cost a few pounds to produce.

The one thing we would not remove from the workflow

This was a small practical test rather than a laboratory study, using one platform and one model, with a single short commercial script in four languages. We can't conclude that Arabic AI VO is universally stronger than Greek, or that human-directed tags will always beat Auto Tag.

But if the finished VO matters, native-language review is essential, and ideally that review comes from someone who understands voice-over rather than simply somebody who happens to speak the language.

Some problems are audible to anyone, like the very British accent that suddenly appeared in the Greek VO. Others, like the issues caused by missing vowel marks in the Arabic or the unnatural break in the Mandarin phrase, are things only a native reviewer would catch. Crucially, only a native ear can judge whether the tone and voice are right for the product in each market. 

A good reviewer will tell you whether the AI needs more direction or the script itself needs to be changed, and help you make those changes and listen again. And, as Arabic demonstrated rather beautifully, sometimes they will tell you that the best thing you could have done was leave a very good first take alone.

Brightlines offers multilingual voice-over with both professional human voice artists and AI voices. If you’re not sure which suits your project, get in touch and we’ll talk you through the pros and cons of each, or visit our voice-over services page to find out more.

Julie Fry Headshot Hi Res
Published by Julie Fry
Senior Project Manager