Chapter 104. Localization: Translate the Action, the Status, and the Cultural Interface#

Localization is not swapping the subtitles. Forms of address, politeness levels, numbers, legal systems, lip sync, jokes, musical conventions, graphic layout and platform rules may all change. Backlit Takeover plans English, Spanish and Southeast Asian versions, so it builds a neutral international master first and adapts regionally afterwards.

104.1 The localization bible, semantic units, and separate voice tracks#

The locale profile.

locale_profile:
  locale: en_US
  character_names: transliterated_with_consistency
  company_structure: adapt_explanation_not_facts
  currency: retain_CNY_with_context
  date_format: MMM_D_YYYY
  honorific_strategy: role_based
  subtitle_reading_speed_cps: 17
  dubbing_priority: intent_over_literal_length
  legal_review_required: corporate_procedure_terms

Facts do not change by region; the interface that expresses them can. If 480 million yuan is converted into dollars, the exchange rate and the story significance must be stated — usually the yuan figure is retained and natural dialogue supplies the sense of scale.

Semantic units and character voices.

The line registry stores the source intent, the action, the subtext, the information that may be cut and the duration — not just the original sentence. A translator first understands that "the meeting continues" is a seizure of procedural control, then chooses a short, high-status expression in the target language.

Each character gets a target-language voice bible: range, tempo, formality, interruption habits and the pronunciation of proper nouns. The English version does not recast a character with a completely different age or temperament.

Subtitles and dubbing on separate tracks.

Subtitles can stay closer to the original meaning; dubbing has to fit lip movement, breath and duration. They are two products from one source, and a subtitle text cannot simply be handed to a voice actor. A dubbing rewrite keeps intent and facts, and deleting a key number or legal relationship requires approval.

When a translation runs long, first change the syntax, split it across two turns, or let the picture carry part of it — do not keep accelerating the delivery. If sync is still impossible, adjust coverage or mouth movement rather than sacrificing intelligibility.

104.2 Graphics, culture and institutions, music and sound#

Graphics and UI.

Contracts, chats, news and phones are re-rendered from field-level facts. Templates reserve room for different language lengths and support wrapping and font changes. A compact vertical Chinese headline may overflow in English, so recompose rather than shrinking type below legibility.

Font licences cover the territory and embedding, with complete character sets. OCR rules are configured for the target language and its number formats.

Cultural and institutional interfaces.

Board procedure, family forms of address and commercial behavior may need light explanation, but a story about a Chinese company should not be forced into US corporate structure. The choice is to keep the cultural specifics and supply context, or to make a full adaptation and revalidate every fact. Mixing two institutional systems is not an option.

Jokes, humiliation and relationship boundaries get a cultural sensitivity review. Changing one form of address can change intimacy and power, so it is written into the relationship state.

Music and sound.

Thematic motifs are kept in principle; lyrics and culturally specific instruments need their own rights and adaptation. Dialogue ducking is rebuilt for each dub rather than layered onto the original mix. Ambience and Foley can be reused, but spoken broadcasts and background conversation must be replaced or pushed below intelligibility.

Loudness, subtitle safe areas and synthetic-content disclosure are re-accepted per platform.

104.3 Localization QC and a transplanted mistake#

Six layers of QC.

Language review, story fact, lip sync and timing, culture, technical specification and rights stay separate. Back-translation catches shifts in meaning only; it cannot tell you whether the target language sounds natural. At least one native-speaking reviewer watches complete episodes at normal speed under phone conditions.

Terminology consistency is checked automatically, with exact checks on names, companies, amounts and dates. Humans handle status, humor, subtext and performance.

A transplanted mistake.

The first English version rendered Lin Xia's line about her standing word for word, which came out abstract, and testers did not understand what the document did. The revision — "Your answer is in the file" — is more explicit and still short, with the legal weight carried by the action of handing it over and the confirmation that follows. The wording is not literal; the action is preserved.

Another version automatically converted 480 million yuan into dollars at the current rate. Weeks later the rate had moved and the graphics no longer agreed with the dialogue. In the end the yuan figure stayed as the fact.

104.4 Repair, acceptance, and a localization workshop#

Fault tree.

Natural subtitles and wooden dubbing: one text was asked to do two jobs. Characters feel recast: the target voice was never locked. UI overflow: templates were designed for one language. Wrong institutional explanation: the localization mixed two jurisdictions. Music burying the dub: it was never remixed. Numbers changing between versions: exchange rates and facts were not governed.

SOP, checklist and deliverables.

Build the locale profile. Translate semantic units. Lock the target voices. Produce subtitles and dubs separately. Re-render graphics. Review culture and institutions. Remix the sound. Run six-layer QC. Export per platform. Maintain the invalidation chain from source changes.

• The translation preserves each character's action and status.
• Subtitles and dubbing are produced separately.
• Graphics are re-rendered from structured facts.
• Native-speaker review covers whole episodes.
• A source version change invalidates every language.

Exercise: localize a 45-second boardroom scene into two languages, delivering locale_profiles/, term_base.csv, subtitle_scripts/, dub_scripts/, graphic_localizations/ and localization_qc.md.

The localization workshop.

Choose a scene containing forms of address, an amount, procedural terminology, an interruption and a phone graphic. The translator writes a meaning-for-meaning pass first, then produces the subtitle script and the dub script separately, noting the reason in action terms wherever they depart from a literal rendering. After a native-speaker table read, record whether status changed, whether a joke disappeared, whether the numbers sound natural, and whether line endings can support the performance.

Lay the target-language audio back over the original picture without lip sync at first, checking the stress on the ideas and the breathing, then decide which shots need local mouth work and which can use reactions and off-screen voice. The graphics team re-flows the UI from fields, deliberately testing the longest name, the longest company name and different date formats. Any version that shrinks type below the device threshold fails.

Finally, change one locked line in the source and verify that subtitles, dubs, graphics, the term base and the release packages all invalidate correctly. If a language never enters the invalidation queue, the version chain is incomplete and you should not add more territories.

The localization evidence pack keeps, per language, the source semantic units, subtitle scripts, dub scripts, voice versions, terminology decisions, graphic fields, native-speaker review and platform technical results. Every manual exception notes which episodes it applies to and when it expires. Do not deliver only the final audio tracks — otherwise the next time a character's name changes, nobody can tell which assets need remaking.

A note on sources#

This chapter describes a localization pipeline rather than any language pair's style guide. What transfers is translating semantic units instead of sentences, keeping subtitle and dub as separate products, re-rendering graphics from facts, and making a source change invalidate every downstream language.