Specify what the audio reference is allowed to contribute. Reusing the recording, repeating its words and borrowing broad performance traits are three different tasks. If you only want a quiet, measured delivery for new dialogue, the sample's old sentence should not enter the target script merely because it is audible in the attachment.
Assign one explicit reference responsibility
The official H3 full-reference guide separates direct reuse from references to timbre, rhythm, emotion or delivery. That distinction is specific documented prompting guidance, not proof that a website supports voice cloning or that any supplied voice can be reproduced faithfully.
Confirm authorization for the intended voice material and use. Then label the reference's responsibility in ordinary words: exact approved soundtrack reuse, exact spoken-content reperformance, or selected delivery characteristics with new approved words. When several responsibilities coexist, name the relevant interval or layer instead of calling the whole attachment “the voice.”
In an invented example, the reference says “The package arrived yesterday,” while the target line is “The workshop opens at nine.” The new brief can retain a restrained pace and clear consonants without preserving the package sentence. Do not place both sentences in a dialogue block unless both are intentionally required.
Keep a content inheritance ledger
| Reference feature | Intended treatment |
|---|---|
| Broad speaking pace | May guide the new performance |
| Old package sentence | Must not appear in the target |
| Background cafe noise | Excluded unless separately requested |
| Target workshop sentence | Sole approved new dialogue |
This ledger also catches accidental ambience transfer. A voice recording made in a cafe may include dishes and music; those sounds are not automatically part of a new outdoor scene. Identify them separately if preserving the complete waveform is truly the task. A prompt cannot separate a mixed recording perfectly just because its layers are named.
Review the words before resemblance
Read the target script without the sample playing. Count its intended utterances and search the working brief for memorable phrases from the old recording. Then listen to the returned media from start to finish. Check for a stray opening word, repeated sample phrase or old sentence added after the approved ending.
A useful proposed acceptance sheet has columns for target words, audible words, speaker and inherited material. Mark uncertain speech for manual review instead of declaring it correct from a transcription alone. If the content is wrong, correct that responsibility before debating subtle voice similarity. This article reports no generation experiment or matching score.
In an Imagild film brief, describe the audio attachment's purpose alongside the actual supported reference label. Do not assume internal H3 audio tags are accepted directly by the user interface. Keep the approved text available for comparison, and decline an output that copies unauthorized or unrequested sample dialogue even if its tone seems convincing.