The brief arrived on a Tuesday: a 90-second product film, already finished in Russian, now needed in 18 languages within ten days. Same presenter, same camera moves, same pacing, and — the client added — the mouth should match. That last sentence is where most localisation projects stop being a translation job and start being a production job, and it is worth being honest about which parts of it are achievable, which parts are expensive, and which parts should simply be dropped.
We have run this pipeline enough times at KAMOD to know that the translation is the cheap part. The rest — voice casting, timing, on-screen text, rights, and one native reviewer per language — is where the budget and the calendar actually go.
Where localisation breaks, and it is not the translation
A machine translation of a 90-second script takes minutes and costs almost nothing. What breaks first is not meaning but register. A phrase that reads as warm and direct in Russian can land as abrupt in German, over-familiar in Japanese, and oddly formal in Spanish for Latin America. Uzbek and Kazakh share a great deal of vocabulary and almost none of the same rhythm; a script read at the same tempo sounds rushed in one and sluggish in the other.
The second thing that breaks is reference. Idioms, sports metaphors, currency, and the way a price is spoken out loud all need rewriting rather than translating. A line like "cheaper than a cup of coffee" is fine in most of Europe and weak in a market where the reference price is different. Localisers should be allowed to rewrite the sentence, not just the words, and the client should sign off on a meaning-not-wording rule at the start. Without that rule, every reviewer sends back style notes instead of errors, and the project loops.
Casting the voice: one voice or eighteen
There are three viable approaches and they cost different amounts.
The first is a single synthetic voice reading all 18 languages. Modern multilingual voice models can do this convincingly in six or seven major languages and acceptably in the rest. The advantage is consistency: the brand sounds the same everywhere, and you pay for one casting decision. The disadvantage is that a voice which is neutral in English can sound like a newsreader in Turkish and slightly uncanny in Polish.
The second is a different synthetic voice per language, chosen from a shortlist of three or four candidates each. This costs more studio time but the result sounds native, because it is. A voice that grew up in the language gets the stress patterns right without being told.
The third is a human voice actor per language for the hero markets and synthetic for the long tail. This is what most serious campaigns end up doing: four or five languages recorded by people, the remaining thirteen generated, all mixed to the same loudness spec so the difference is not audible between versions.
Ask any studio which of the three they propose before the quote is signed. A quote that does not name the approach is a quote that will change.
Lip-sync: what is achievable and what is not
Lip-sync in AI video splits into two very different problems.
On a talking-head shot, where the face occupies a third of the frame and the camera is locked, modern lip-sync models do well. They retime the mouth to new phonemes, and at normal viewing distance on a phone the result holds up. The failure modes are specific: teeth, tongue, and vowel shapes that do not exist in the source language. Arabic and Hebrew need throat and guttural shapes that a model trained mostly on English will approximate badly. When the shot is a close-up, plan for a manual frame-by-frame review of the first ten seconds before approving the rest.
On a wide shot, a profile, a moving camera, or a presenter whose mouth is partly turned away, lip-sync is not the right tool. Here the honest answer is a cutaway, a B-roll insert, or a voice-over version where the presenter is not speaking on camera. Forcing lip-sync onto a three-quarter turn produces the rubbery, uncanny result that makes viewers distrust the whole film.
There is a third case worth naming: a presenter filmed speaking one language, localised into a language with a very different mouth rhythm. Even with good lip-sync, the head movement and blink rate were timed to the original. If the new audio runs 20 percent longer, the presenter appears to hold a pose too long. The fix is to allow a slight speed change in the audio, usually up to 8 percent, before touching the picture.
On-screen text is a second film
Any text burned into the frame — a price, a product name, a call to action, a legal line — has to be redone for every language, and it is not a find-and-replace. Text sits in a designed composition: it has a size chosen to fit a specific phrase, a line break chosen for a specific reading rhythm, and a safe area chosen around a specific crop.
Three practical rules save weeks. Keep a layered master with text on its own layer, so every language can be re-typeset without regenerating the shot. Lock the text style to a typeface that covers Cyrillic, Latin with diacritics, Arabic, and Devanagari, or accept that you will need two type systems. And plan for expansion: German and Turkish commonly run 20 to 35 percent longer than the English equivalent, which means a headline that fits on one line will not, and the composition has to be designed with that space in reserve.
Right-to-left languages are their own pass. Arabic and Hebrew do not just mirror the text, they mirror the reading order of a layout, and a call to action placed at the right edge of the frame is now at the wrong end of the eye path.
Timing, subtitles, and the three delivery formats
Every project should ship three formats, because they serve different platforms.
Dubbed audio over the original picture is the primary deliverable for social and for a website hero. Subtitled versions are what a large share of viewers will actually use, especially on mobile with sound off, and they should be delivered as burned-in captions for social plus a separate SRT for platforms that support toggling. The third is a localised voice-over mix, where the original audio is ducked rather than removed — useful for markets where the audience expects to hear the original speaker.
Then there is the audio bed. Music and sound effects should be delivered as a separate stem, because a rights-cleared track in one territory is often not cleared in another, and swapping music per language is much cheaper than remixing the whole film. Loudness should be normalised to a single spec across all versions, typically minus 14 LUFS for online delivery, so that no language version arrives quieter than the others on a platform.
The QC pass that catches what the model cannot
No synthetic or human voice should ship without a native speaker listening to the finished file, not the script. The reviewer needs to hear the audio against the picture, at normal speed, on a phone speaker. The notes worth collecting are narrow: mispronounced product names, numbers read the wrong way, a sentence that lands as rude, and any place where the audio and the mouth visibly disagree.
Give the reviewer a glossary at the start: how the brand name is pronounced, whether the price is read in full, whether the presenter uses formal or informal address. Uzbek has a real distinction between siz and sen, Russian between vy and ty, and getting it wrong in a 30-second ad is the kind of error viewers comment on.
Build the multilingual version of the film as a batch, but review it as individual films. A single build that produces 18 files in an afternoon is efficient; approving all 18 in that same afternoon is how a mistake ships in fourteen markets at once.
If you are planning a localised campaign, send us the master script and the platform list before the picture is locked. Half the cost of dubbing is decided by decisions made during the shoot, and KAMOD works with brands in Tashkent and internationally on exactly this sequencing — story, then picture, then languages, in that order, because reversing it roughly doubles the price.