The most expensive localization error isn’t a mistranslated word. It’s rewriting the script, recording the voice, shipping the captions, and only then discovering the timeline no longer fits. That’s rarely a skill problem — it’s a missing checklist.
Here’s one you can paste straight into your project tool. Grouped by stage, every item with a pass criterion. It applies to any “one source script → many languages” workflow.
Why localization QA needs its own gate
| Common shortcut | What breaks |
|---|---|
| Translate, then voice | Translations are written for the eye, not the mouth |
| Voice, then drop in captions | Captions drift out of sync with actual delivery |
| One language passes, ship them all | Languages differ in length; timelines aren’t shareable |
| Spell-check only | Pronunciation, stress, and cultural tone go unchecked |
The core idea: translation, voice, and captions are three separate assets, each verified separately. One passing says nothing about the others.
Stage 1: The translated script (before voicing)
Cheapest place to fix anything, so catch most of it here.
| Check | Pass criterion | Failure looks like |
|---|---|---|
| Speakability | Read it aloud without stumbling | 25+ syllables per sentence, three stacked clauses |
| Natural order | A native speaker hears no “translationese” | English passive voice carried into Chinese intact |
| Term consistency | One term, one rendering, throughout | “Cloud” in one section, “cloud service” in another |
| Numbers and units | Localized (units, decimal marks, date format) | Imperial units left in a metric market |
| Brand names | Kept or transliterated per brand rules | Brand name invented by the translator |
| Idioms | Replaced with the local equivalent | Literal idiom that reverses the meaning |
| Register | Matches the source’s formality | Friendly source, bureaucratic translation |
| Sensitive terms | Market taboos and political sensitivities removed | Religion, history, or place names that land badly |
Practical ask: have the translator deliver a length estimate per section, expressed as a percentage over the source. That number decides whether the script needs cutting later.
Stage 2: Script adaptation (before voicing)
The same line can differ 30% in length across languages. Rewrite for rhythm per language instead of forcing it in.
| Check | How |
|---|---|
| Paragraph length | 2–3 sentences, so sections can be regenerated alone |
| Pace baseline | Decide the target pace per language up front (table below) |
| Pause markers | Write pause positions and durations into the script |
| Emphasis | Mark which word carries stress; don’t let the model flatten it |
| Number reading | State whether a number is spoken as digits or words |
| Language | Typical target pace | Note |
|---|---|---|
| Chinese (Traditional/Simplified) | 190–260 characters/min | Slow end for teaching and health content |
| English | 125–165 wpm | Fast end for ads, slow end for instruction |
| Japanese | 300–380 characters/min | High mora density; sounds faster than it counts |
| Spanish | 150–190 wpm | Many syllables; budget extra time |
| German | 120–150 wpm | Compound words stretch sentences fast |
These are starting points, not standards. Test 30 seconds of your own material — that beats any table.
Stage 3: After voice generation
| Check | Pass criterion |
|---|---|
| Proper-noun pronunciation | Verified against the pronunciation sheet, item by item |
| Heteronyms | Chinese polyphonic characters, Japanese kanji readings, English heteronyms all confirmed |
| Pace | Within ±10% of target |
| Pauses | Sentence, paragraph, and post-instruction pauses match the script marks |
| Stress | Lands on the marked word; no uniform delivery |
| Emotion | Matches source tone — not energized in every language |
| Endings | No truncation, no stray breath or noise |
| Consistency | Same character, no timbre drift across sections |
Item 8 is the one people miss. When a long script is generated in pieces, timbre, pace, and level can shift subtly between sections. Either generate a section in one pass, or lock seed and parameters.
Three segments that always get a human listen
- The first sentence — where viewers decide to stay.
- Every number and unit — the least noticeable errors.
- The last line and CTA — most often clipped or rushed.
Stage 4: Timeline and sync
| Check | Pass criterion |
|---|---|
| Total length | Within ±10% of the original, or picture re-cut |
| Sync points | Every visual cut aligned to a sentence boundary |
| Caption timing | No more than 0.2s early, 0.5s late |
| Breathing room | 0.5–1.0s between sections; no hard butt-joins |
| Music bed | Voice sits 6–8 dB above music |
| End card | Enough time to read the CTA and legal line |
When length drifts past 15%, don’t fix it with speed. Cut content first, adjust pace second — sped-up voice loses credibility, especially in teaching and health material.
Stage 5: Captions and transcripts
| Check | Pass criterion |
|---|---|
| Line length | ≤18 Chinese characters or ≤42 English characters per line |
| Line count | Two lines maximum at once |
| Minimum duration | At least 1 second per caption |
| Breaks | At semantic boundaries, never mid-word |
| Punctuation | Follows local convention (full-width for Chinese) |
| Matches the voice | Compared line by line against the final script |
| Proper nouns | Same rendering as the spoken version |
| Legibility | Readable on a phone; white text backed by shadow or box |
The classic failure: captions describe a different version than the audio. It happens when captions are generated from the translation while the voice used the adapted script. Captions must come from the final voiced script.
Stage 6: Cultural and market fit
| Check | Why |
|---|---|
| Address and honorifics | Japanese, Korean, and Southeast Asian markets need the right politeness level |
| Humor | Jokes rarely survive translation; cut rather than explain |
| Examples and place names | Swap for something the market recognizes |
| Color and symbols | Some colors and gestures carry negative meaning locally |
| Regulatory wording | Required phrasing for claims and disclaimers |
| Currency and pricing | Local currency and format; re-quote where needed |
| Seasonal timing | Avoid launch dates that collide with local holidays |
There’s no universal answer here — it needs a local reviewer. Native speaker ≠ local reviewer. Correct language isn’t the same as correct market sense.
Stage 7: Final pre-launch gate
- [ ] Every language has its own final script and audio
- [ ] Pronunciation sheet updated and archived
- [ ] Captions generated from the final audio and checked
- [ ] Timeline aligned to every visual cut
- [ ] Loudness and peaks within platform spec
- [ ] Length deltas per language recorded
- [ ] Legal lines and disclosures adapted per market
- [ ] Naming and version scheme applied (below)
- [ ] Old versions marked or pulled
- [ ] Sign-off recorded — one language reviewer, one market reviewer
Versioning and naming
A scheme that survives multiple languages and revisions:
<project>-<language>-<version>-<date>
e.g. spring-launch-zh-TW-v3-20261115
spring-launch-en-US-v3-20261115
spring-launch-es-MX-v2-20261110
Supporting rules:
- Use full locale codes (
zh-TW,en-US,es-MX), never barezhoren. - Bump the version on every script change; never overwrite.
- Keep script, audio, captions, and generation settings as one bundle.
- Log which page or channel each version shipped to.
What automation can and can’t do
| Automatable | Still human |
|---|---|
| Naming and version checks | Pronunciation accuracy |
| Caption length and duration | Whether the tone is right |
| Length delta calculation | Cultural nuance |
| Loudness and peak detection | Whether it sounds translated |
| Missing-section comparison | Legal and market compliance |
Automation clears the low-level errors so people spend their time on judgment. That’s the real value of AI voiceover: not removing people, but putting them where judgment happens.
Batch every language with text-to-speech, keep a consistent brand voice with voice cloning, and see usage and plans on the pricing page.
FAQ
Can we voice first and fix captions later? Yes, but the order must be voice-final, then captions. Do it backwards and the captions get redone.
What if language lengths differ a lot? Re-cut picture and adjust spacing first, cut content second, adjust pace last. Anything past 10% faster is audible as strain.
Does every language need a native reviewer? Yes — at minimum one native speaker plus one in-market reviewer. High-risk content (medical, financial, regulatory) adds specialist review.
Can AI voice stay perfectly consistent? Within a section and fixed parameters, closely. Across sections or sessions, lock parameters and spot-check by ear.
How many languages at once? Ship two, get the process and checklist working, then expand. Six languages with no checklist multiplies errors, not reach.
How detailed should the checklist be? Detailed enough that a different person gets the same result. If it just says “check translation quality,” it’s a wish, not a checklist.
Turn the checklist into an advantage
Localization quality doesn’t come from one big review. It comes from running the same checklist every time. When every language and version goes through the same seven stages, errors become predictable and fixable items instead of post-launch surprises.
If your content touches medical or patient education topics, tone and compliance get stricter — see AI voiceover for patient education videos. For the wider multilingual production flow, read multilingual product video voiceover.


