Producing a second-language video involves more than converting speech. Source footage may contain baked-in captions, diagrams, navigation labels, product names or visual jokes that do not translate through a dubbing engine. A production-ready multilingual master has to reconcile all layers of meaning while preserving accessibility and original project files.
Freeze the source script
Before any translation, assemble the approved source transcript with speaker IDs, exact numbers, source-language names and a glossary. Correct accidental speech-to-text errors. For safety-sensitive or regulated material, identify which passages must be reviewed by a specialist rather than a general translator. Keep the source video and original subtitles as a baseline, never overwrite them with the new version.
Separate three media layers
First, the spoken dialogue can be dubbed or revoiced. Second, the visible speaker’s mouth may optionally be modified. Third, on-screen text and captions must be localized separately. HeyGen’s video-translation help explicitly states that spoken-audio translation does not replace on-screen graphics or hard-coded captions. Synthesia’s dubbing flow similarly provides language and duration choices, but the editor must still verify timing and written material.
Create a localized shot spreadsheet with timecode, original line, translated line, overlay replacement and reviewer status. If a translated phrase is longer, decide whether to retime the slide, adjust the script or add a subtitle line break. Do not speed an instructional sequence so aggressively that technical steps become impossible to follow.
Accessible caption timing
Prefer separate SRT or WebVTT assets for platforms that support selectable captions. These can be corrected without rerendering a video. Burned-in captions may be useful for short feeds that autoplay silently, but they are less flexible and can conflict with platform accessibility tracks. Apply a style guide for line length, safe areas, contrast and speaker identification. Review the final image after cropping to vertical formats: readable desktop text can be cut off on a phone.
Cultural and regional review
Translate user-interface wording against the localized product, not a machine literal translation. Verify currencies, dates, country-specific terms and measurement units. Where a narrator pronounces a foreign personal name or a trademark, obtain an approved pronunciation; don’t let a voice-clone approximation rewrite the name. Use a native-language review pass on both audio and subtitles.
Versioned delivery
Each locale receives its own export manifest: language code, audio mix, subtitle file, caption type, model/version used for any synthetic performance, consent scope and sign-off. Keep one source-of-truth edit project and generate local variants from it. AI 05 manages underlying transcription or voice-product comparisons; this guide governs their integration with video editing and publication.
Worked production example
A seven-minute lesson includes subtitles, a branded slide, a safety warning and spoken explanation. Before dubbing, extract the warning from the graphic layer so it can be replaced in each language. The translator checks units and names, the editor checks that slide transitions follow the new narration, and an accessibility reviewer checks SRT order and timing. The final master must be viewed on both a desktop screen and a phone. A video can contain correct translated speech while still misleading viewers through untranslated on-screen text.