For a bilingual scene, assign the spoken language to each approved utterance instead of setting one vague language for the whole film. Keep the exact dialogue, displayed text and interface language as separate objects. An English production description does not require the characters to speak English, and a Chinese interface does not establish what language the returned audio uses.
Preserve the words before adding language metadata
The MiniMax H3 base guide describes language-labelled dialogue while preserving user-provided speech. Its rewrite instructions distinguish descriptive prose from the actual utterance. This is an H3 convention; it is not a claim that a language label guarantees pronunciation, dialect or support in every deployment.
Create a line ledger with speaker, exact words, intended language and whether a translation is only explanatory. A gloss for the editor must not become another line to be spoken. For names and short borrowed phrases, ask what the creator actually intends to hear rather than inferring from the alphabet alone.
Review switches at real conversational boundaries
In a fictional workshop scene, a host greets a visitor in English, gives an explanation in Chinese and closes with an approved English phrase. The language changes are part of that particular conversation, not a request to translate the entire script into two full parallel versions.
| Ledger field | Purpose |
|---|---|
| Exact approved line | Defines the content to preserve |
| Intended spoken language | Defines how that line should be performed |
| Editor's translation | Helps review meaning; not additional speech |
| Speaker identity | Keeps the voice role stable across switches |
A genuinely mixed-language sentence needs a deliberate choice. Do not insert arbitrary new pauses just to force every borrowed word into its own utterance. Follow the actual supported representation and retain natural phrasing. If the integration cannot express the needed switch reliably, plan separate authorized performances or simplify with the owner's approval.
Compare text preservation with audible performance
Before generation, compare the rewritten lines with the approved text for omissions, accidental translation and changed names. Internal H3 formatting must retain the complete required request structure; isolated example tags do not form a complete prompt and are not the website's attachment syntax.
An internal fragment such as <d>[English] Welcome back.</d> illustrates a language-labelled utterance, not a complete submission. Its descriptive context can remain English while a separate approved Chinese line retains its Chinese words. Do not put the editor's translation beside it as another spoken line.
After generation, have a competent speaker review actual audio for language, words, pronunciation and natural boundaries. A transcript can miss an accent error or normalize an unintended translation. Check whether the character remains the same speaker after switching language, rather than assuming a new language requires a new vocal identity.
Put the bilingual line ledger beside an Imagild film brief using the site's supported inputs. Accept when each approved line is actually spoken as intended and explanatory translations remain outside the soundtrack. This guide provides a planning and verification method, not a completed multilingual benchmark. For a language or dialect that the chosen route does not support, state the limit and choose another approved production path rather than relabelling an incorrect output.