A translated voice can sound assured while saying the wrong thing. It can also say the right thing while the captions show an older script. A dependable localisation workflow therefore needs more than a successful audio render. It needs agreement between the intended meaning, the words actually spoken and the text viewers read.
We call these three scripts the meaning script, the performance script and the caption script. These are editorial working labels, not a technical standard. Their purpose is to make revisions traceable when one sentence moves through translation, synthesis and editing.
The meaning script is the reference
Start with an approved source and a reviewed translation. Identify names, product terms, numbers, units and any phrase whose meaning depends on context. Put the approved interpretation beside those items. If a number changes, that change should reach every version of the programme rather than only the latest voice track.
Give a reviewer enough context to judge the translation. A short sentence displayed alone can conceal whether “it” refers to a product, a person or an action. Supply the relevant scene and intended audience. A fluent sentence is not automatically the right sentence for that scene.
The performance script controls what is heard
A speech system may need a particular spelling or supported pronunciation instruction to produce an intended name. Keep such interventions in a separate performance version. Do not silently replace the approved written name everywhere because a phonetic spelling produced better audio.
Review short, difficult passages before producing the entire programme. Include abbreviations, mixed-language terms, numbers and transitions between sentences. Listen to the actual output, not only the text sent to the service. Keep the approved audio file with the exact script revision and settings used to create it.
Voice identity introduces another decision. ElevenLabs’ Professional Voice Cloning documentation describes a voice-verification step and limits that feature to cloning one’s own voice. That is a product-specific condition, not a general rule for every voice service. Check the rules and permissions for the actual workflow rather than assuming that uploading a recording is enough. ElevenLabs documentation.
The caption script follows the final sound
W3C distinguishes captions, which convey speech and necessary non-speech audio information, from translated subtitles in its own terminology. It also warns that automatic captions require accuracy review. A speech-recognition pass is therefore a starting point, not the final editorial check. W3C WAI guidance on captions and subtitles.
Build or retime captions against the approved audio. Check names and numbers against the meaning script, then check the timing against the performance. If a line was shortened for pacing, confirm that the change still communicates the intended information. Do not copy a phonetic helper spelling into the visible captions.
Watch the complete scene with captions on. They must coexist with labels, diagrams and changes of shot. A perfectly transcribed line can still arrive too early, cover useful information or remain on screen after the speaker has moved to a new subject.
A revision example
Imagine an instruction that changes from “connect the second cable” to “connect the marked cable.” This is an illustrative production scenario. The meaning changes even if the new sentence is the same length. Regenerate or edit the relevant audio, update the caption text and inspect the image for a conflicting label. A timing match alone would miss the inconsistency.
Mark the previous assets superseded rather than overwriting them without a record. The editor needs to know which soundtrack and caption file belong together. Use one approved programme version as the unit of release, with its component files listed.
Acceptance belongs to a listener
Use a reviewer who can assess the target language and the subject matter at the level the programme requires. Ask them to listen without looking at the written translation for one pass, then compare against it for another. The first reveals unclear delivery; the second exposes omissions and altered meaning.
Finish with the exported video or audio package, not the service’s preview player. Verify the right language, the right voice, the approved text and the corresponding caption file. A synthetic voice can reduce recording friction, but it does not remove the editorial responsibility to know what the audience actually receives.