Machine Translation Limits: What Breaks in Dubbed Scripts

Machine translation has become the backbone of fast dubbing workflows. It handles millions of word pairs, catches obvious grammar, and turns your script into 50 languages in seconds. But it has sharp limits, and those limits show up in ways that wreck the feeling of a video.

When a translation misses an idiom, flattens humor, mixes up units, or picks the wrong formality level, viewers hear it. They might not consciously know what went wrong, but they know the dub sounds off. The good news: most translation failures are predictable, and catching them before you publish is cheaper than reshooting.

This guide walks you through the most common translation pitfalls, shows you what they look like in real scripts, and gives you a concrete checklist to catch them before your video goes live.

Idioms and Expressions: Word-by-Word Breaks Context

Machine translation reads word by word and builds meaning from vocabulary alone. It has no knowledge of cultural context or historical usage. "It's raining cats and dogs" becomes something like "animals are falling from the sky" in most languages, because the statistical model never learned that English speakers use this phrase to mean heavy rain.

Idioms rarely translate one to one. Spanish has "llueve a cántaros" (it rains pitchers), French says "il pleut des cordes" (it rains ropes), and German uses "es regnet Bindfäden" (it rains twine). A translator who knows the language picks the local idiom. A machine picks the literal path every time.

This matters most in comedy and casual conversation, but it also breaks training videos and customer testimonials. A CEO saying "we broke the ice with our new market" becomes "we shattered frozen water" unless you catch it. A support agent saying "I'm not going to sugarcoat this" might become "I'm not going to put sugar on this" in literal translation, sounding absurd.

The symptom viewers notice is a phrase that sounds stiff or makes no sense in context. When they hear something unnatural, they lose trust in the video, even if they cannot name the problem.

Worked Example: E-commerce Testimonial

An American customer in a product demo says, "This tool is the real deal. No nonsense, no beating around the bush." Machine translation to Spanish produces: "Esta herramienta es la cosa real. Sin sinsentido, sin golpear alrededor del arbusto." That third sentence sounds like someone is literally whacking a bush, and viewers in Spain will think the translator failed.

A native speaker would write: "Esta herramienta es de verdad. Sin complicaciones, directo al grano." That preserves the meaning and sounds natural.

Humor and Wordplay: Jokes Do Not Travel

Translation breaks jokes because jokes depend on sound, multiple meanings, or cultural reference. A pun on the word "bank" (financial institution or river embankment) works in English but has no mirror in languages where the word for financial institution sounds nothing like the riverbank.

Wordplay is almost never salvageable through direct translation. A machine might flag that something is funny (repeated syllables, unusual word pairing), but it cannot rewrite the humor for a new language. That requires a human who knows both cultures and can rebuild the joke from intent instead of words.

Comedy dubbing usually needs a separate pass from a native speaker who rewrites the script specifically for the target language, trading exact words for the same laugh. This is the single most expensive fix in any dubbing project. If humor is central to your video (sketches, comedy reviews, satirical content), budget extra time for this rewrite.

Worked Example: Tech Tutorial with a Joke

A software tutorial says, "Debugging is like being a detective, except the criminal is your own code." The machine might translate that literally, but it misses the humor. In Japanese, this might become "debugging resembles detective work, except the criminal has written code." The setup is there but the punch is gone.

Restoring it requires someone who can rewrite for Japanese audiences: "デバッグは推理ゲーム。犯人は常に、自分の書いたコードだ。" That structure works better in Japanese comedy, where the reveal comes at the end.

Units, Measurements, and Currency: Context Collapse

Machine translation often leaves measurements and currency untouched or converts them wrongly. A script that says "5 pounds" might stay as pounds in a language where locals think in kilograms, or the translation might say "5 kilo" without checking whether the original meant weight or sterling.

Temperature is famously bad. A prediction of "72 degrees" in Fahrenheit is pleasant and comfortable. In Celsius, it is literally boiling and deadly. Translation systems rarely notice this context switch and simply convert the number or paste the original, leaving viewers confused or horrified.

Currency names need localization too. Saying "dollar" in a market that uses euros feels foreign and confusing, especially in older videos where the exchange rate has shifted. A script quoting "$50 per month" becomes meaningless if you do not convert to local currency or remove the symbol entirely.

Size and weight standards vary too. A "large" shirt in the US is not a large shirt in Asia. A "10-pound weight" in a fitness video needs conversion and context for viewers using kilograms. Native speaker reviewers catch these instantly because they think in their local units.

Worked Example: Pricing Video

A sales video says, "You get 15 pounds of coffee beans for $35 per month." In the UK, this should be kilograms and pounds sterling, not pounds of weight and dollars. A German audience reads "Pfund" (currency) and "35 Dollar" and feels the pricing is for an American market, not for them, and may not even check if it is available in their country.

The fix is one sentence: "Receive 7 kilos of coffee beans for 32 euros per month." That one change makes it feel local and relevant.

Names, Brands, and Proper Nouns: When Systems Over-Translate

Good translation systems leave proper nouns alone. A person named "Grace" should stay "Grace" in every language. A brand name like "Coca-Cola" does not translate.

But some systems do translate names ("Gracia" instead of "Grace") or misidentify what is a proper noun versus a common noun. In Japanese, many names have meaning and a system might translate the meaning rather than preserve the name. A person named "Yuki" (meaning snow) might get translated to "Snow" in English, losing the name entirely. Product names sometimes get auto-translated too, which is a disaster for brand videos or sponsor content.

Always check that every name in your dubbed script matches the original. This is a quick scan that catches obvious mistakes and prevents brand damage.

Formality Levels and Tone: Flatness Kills Character

Language formality has no single dial. English is relatively flat: "you" works in formal and casual settings. Spanish has "tú" and "usted". German has "du" and "Sie". French has "tu" and "vous". Many Asian languages have five or more levels of politeness, each carrying social meaning.

A machine might pick a middle formality level that sounds wrong for your context. A CEO presenting at a formal conference needs formal language. A teenager in a comedy sketch needs casual speech. Machines often flatten both to a neutral middle ground that sounds like no one.

This is particularly important for interviews, training content, testimonials, and anything with character voices. Each speaker should have a consistent tone that matches their role. The right formality level sells that character and builds trust.

Worked Example: Executive Interview

An executive says, "We're excited to announce this new partnership." In Spanish, a machine might generate "Estamos emocionados por anunciar esta nueva asociación," which is neutral. But the executive needs gravitas. A native speaker would write "Nos complace presentar esta nueva asociación estratégica," which sounds more official and appropriate to the role.

Slang, Colloquialisms, and Regional Speech: Hidden Errors

Slang and regional colloquialisms almost never translate. "That's sick" (meaning good) becomes "That is diseased" if a machine reads it literally. "This deal is fire" becomes "This agreement is combusting." Australian "G'day mate" has no direct equivalent in most languages.

Regional speech patterns matter too. British English "brilliant" is not the same as American "awesome," and machines do not preserve these distinctions well. A script recorded by a British speaker but translated to American formality levels will sound like the character changed.

Generational slang is equally difficult. Gen Z slang like "it's giving..." or "no cap" has no equivalent in most target languages. Machines cannot infer that "cap" means "lie" or that "it's giving" means "it conveys the impression of." They will either paste the English slang (which sounds wrong to non-English speakers) or translate it literally (which sounds absurd).

The broader problem is that slang changes every few months. A script that sounds contemporary today might sound dated in three months. Machine translation cannot track these shifts. A native speaker reviewing the script will immediately flag language that feels off for their audience's current vernacular.

Worked Example: Product Review Video

A British reviewer says, "This app is brilliant, yeah? Really sound design." Machine translation to German might produce something literal that loses the British casual tone. A native speaker would adjust to German equivalents: "Diese App ist hervorragend, nicht wahr? Wirklich gutes Design." That maintains the approving tone without trying to preserve the Britishness.

Similarly, if a social media influencer uses Gen Z slang like "this product is no cap," a machine will either paste "no cap" into the German translation (confusing viewers) or translate it to something like "this product has no hat," which is nonsensical. A native speaker knows to rewrite it as "Dieses Produkt ist wirklich zu empfehlen" (this product is really worth recommending) to maintain the authentic endorsement tone.

Common Translation Mistakes and How to Fix Them

Below is a table of the most frequent errors, what they look like, and how to catch them:

Red Flags to Catch Before Publishing

Before you publish a dubbed video, read through the translated script yourself, or better yet, have a native speaker scan it for these specific things:

• Any phrase that makes you pause or sounds stiff or unnatural
• Numbers, dates, or currency that look wrong for the target market
• Names that were changed or sound unfamiliar compared to the original
• Repeated words or phrases that feel monotonous or awkward
• Tone that does not match the speaker's role (too formal for casual, too casual for serious, wrong cultural register)
• Anything in the original that was a joke, pun, or wordplay
• Temperature, weight, or distance values that use unfamiliar standards
• Words or phrases that feel like they belong to a different region or country

Flag these before voice synthesis runs. Fixing a script takes 15 to 30 minutes; re-voicing takes hours.

The Workflow: When and How to Review

Machine translation handles the mechanical work well. But every dubbed script benefits from at least one pass by someone who knows the target language. That does not have to be expensive or slow.

Here is the most efficient workflow:

1. Generate the translation using machine translation. This takes seconds and gives you a baseline.
2. Read it yourself if you know the language. Take 5 to 10 minutes and flag anything that sounds off.
3. Have a native speaker review it. Ask them to spend 15 to 20 minutes reading the translated script and flag anything that sounds unnatural, wrong for the market, or out of character.
4. Focus their review on the checklist above, not on the entire script. This is faster and more reliable than asking for a complete edit.
5. Batch their feedback into a single revision pass. Do not re-translate; edit the translation to fix only the flagged issues.
6. Run voice synthesis only after the translation review is complete.

This workflow catches most of the problems that make viewers feel like something is off with the dub.

What to Do Next

Start with your next dubbed video. Before you publish, take these concrete steps:

1. After machine translation generates your script, do a quick self-check using the red flags list above.
2. Find a native speaker in your target market. Many freelancers offer 15 to 30 minute script reviews for $20 to $50 per language. Budget this as a standard part of your workflow.
3. Create a small shared document (Google Docs, a simple markdown file) and ask the reviewer to add comments only at lines that need fixes. This keeps feedback focused and prevents scope creep.
4. Edit the translation to fix flagged issues. Keep the edits minimal and focused on the translation itself, not on creative rewrites. This prevents introducing new errors while fixing old ones.
5. Do not re-translate the entire script unless more than 30% of sentences are flagged. A fresh translation often introduces new errors while solving old ones.

For comedy, testimonials, or character-driven content, add 20 to 30 extra minutes for a deeper review or a small rewrite pass. This is where the biggest quality gains happen. A skilled native speaker can preserve humor and character voice in ways machine translation simply cannot match.

The goal is not perfection. It is catching errors that make viewers wonder if something went wrong with the dub. Most viewers will not consciously notice perfect translation, but they will immediately sense when something is off. This review step is your insurance against that moment.

---

🚀 Start Dubbing Your Videos Today

DubLab uses AI to translate your videos into 92+ languages in minutes.

📱 Download for iOS

🌐 Try Free at dublab.app