AI translation has come a long way in the past two years. For ordinary text, there’s nothing left to worry about: it translates a website, product descriptions, and an hour-long recording, and does a decent job. But it reliably fails on one layer nobody watches. Those are the words that were never in a dictionary: your name, your brand name, and your product name.
I ran into this while preparing a foreign-language version of a client’s recordings. The text was translated, the voice was ready, and everything looked done. Before we started generating, we went through nineteen recordings by hand. It turned out that without that pass, sixteen of them would have gone out with an error.
Translation knows the language, but not your name
A translation tool watches grammar and meaning. Proper names and brands fall outside that logic, though, because there’s no correct answer for them that could be derived from rules.
So the result passes any automated check. The text is grammatically correct, the sentence makes sense, and the tool is confident. The error only shows up by ear, and only to the ear of someone who really speaks the target language.
Language trap
An error a translation makes outside the dictionary, in names, brands, and product names, while the grammar stays correct. The tool doesn’t flag it, because it considers the result correct. Only a native speaker catches it, by listening rather than reading.
The costliest translation errors are invisible: they’re the ones the tool marks as done.
Trap one: a surname that changes form in Czech
This is the most treacherous of the three. A Czech surname is spelled differently in each grammatical case, and machine translation simply takes whatever form happens to be in the source sentence.
With the client I mentioned, this would have carried a declined form of her surname, with its Czech case ending, into the English recordings. It affected sixteen of nineteen recordings, the vast majority of the library. An English-speaking listener would have heard a name that makes no sense in their language and remembered it wrong.
The tricky part is how the roles split. To someone who doesn’t speak Czech, the form doesn’t seem odd, because they have nothing to compare it to. And to someone who does speak Czech, it seems completely normal, because that’s how they know it.
The fix cost nothing. The correct form of her name already existed in her own original recording, so we cut it from there and used it. It sounds exactly like her, because it is her. Just like with a voice clone that narrates articles, the best source is often the one you already have at home.
Trap two: a brand name reads differently in another language
A brand name written in Czech often gets read in English in a completely different way than you intended. Same letters, different pronunciation.
On a text-only website, that doesn’t matter, because nobody says the name out loud. But as soon as it gets into a voice recording or a video, your company name starts to sound like a foreign word. The worst case is when it’s read as an existing English word with a different meaning. Then the listener remembers that other thing.
You can test this for free. Have a voice tool read the name in the target language and listen to what comes out. With tools that handle Czech, it’s worth comparing two or three, because each one reads foreign words differently and the differences tend to be big.
Localization
More than translation. Besides the language, you also adapt names, units, date formats, examples, and tone. The goal is a text that feels as if it had been written in that language from the start. All three traps in this article fall under localization.
Trap three: the name in the very first sentence of a recording
The third trap is about organization rather than language. A name or brand name usually sits in the first sentence of every recording, because that’s how recordings naturally open. First you say who you are, and only then do you get to the point.
That means a single error in a name gets copied into every piece in the library. And because it’s always at the beginning, the listener hears it before anything else. Fixing it after generation means going through the whole library again.
A one-minute self-check
Write down three things: your surname, your brand name, and the name of your main product. For each one, answer two questions: how should it be read in the target language, and does it appear in the very first sentence of the recording? Wherever you answer “I don’t know,” you have a trap.
No machine guards this layer
All three traps have one thing in common: they pass the checks, and only a person who has spoken the language since childhood catches them. The ear decides here, not grammar. A foreigner with excellent English will miss a small thing like that.
I know the same problem from localizing an interface into Czech, where a single word on a button gets stuck and nobody notices for months. The same goes for Czech AI features in a platform, where the nuance of the language decides usability more than the feature itself.
It’s exactly the role an AI employee in a business leaves for a human. It does ninety percent of the work and waits for approval on the rest.
So who isn’t this check worth it for? If you’re translating one page with no names or brands, there’s nothing to check, and the tool is enough. But as soon as something goes out under your name, it’s always worth it. The cost of a fix grows with the number of finished pieces, and once the whole library is generated, it’s too late.
The whole process in a nutshell
- List your names, brands, and product names before you start translating.
- For each one, determine the correct form and pronunciation in the target language.
- Check whether they appear in the very first sentence of your recordings.
- Have a native speaker go through the list, not just a translator.
- Only start generating the whole library once the list is approved.
Let’s go back to those nineteen recordings. Nothing that was about to go wrong there was visible in the text, and no tool flagged it. A good translation check hunts for the words where grammar plays no part at all.
