Assign every line to one speaker and specify what the listener does during it. A two-person picture plus a block of dialogue does not reliably establish turn ownership. Review words, voice and visible mouth movement separately. A plausible face is not enough if the wrong person appears to deliver the promise or reveal the clue.
Write a turn list before a sound brief
Record speaker, exact line, listener and purpose. The purpose might be asking a question, refusing an offer or admitting a mistake. Add a visible response only where it helps comprehension. Avoid asking both people to smile, speak and turn at the same instant unless that overlap is intentional and testable.
Treat silence as a real turn. A listening pause can explain who received the line and how they understood it. It also gives an editor a reaction shot if generated speech is unreliable. Do not attribute usable lip sync or multi-voice controls to Imagild without confirming the actual model and interface.
A fictional interrupted vow
In an invented wedding comedy, Mara says, “You handle the wind. I'll hold the roof.” Theo answers, “Together, this time.” The first line gives Mara an action and Theo a task; the answer confirms cooperation. If Theo's mouth delivers the first line, the action ownership becomes ambiguous even if the voices sound distinct.
A reviewable plan shows Mara speaking, Theo listening and then answering. An alternative uses narration over a wide shot and keeps both mouths neutral. That is a different treatment, not a claim that the original dialogue shot was repaired. Choose it openly if the generation cannot sustain the required turns.
Test ownership in three passes
First read the transcript without picture. Confirm who is responsible for each action. Next watch without sound: see which face appears to speak and whether the reaction belongs to the listener. Finally play them together to check that the visible and audible roles agree.
Do not infer the correct voice from appearance or assume that a shared language setting guarantees turn assignment. If a shot changes speaker midway, write the exact point in a revision note. A broad request for more natural conversation may improve expressiveness while preserving the wrong ownership.
Choose a fallback that preserves the story
In the film workspace, review the story and reference roles before creating video. Where supported, revise the affected shot. For a difficult conversation, use separate speaker views, a reaction or a clearly separate voiceover assembled in an external editor. A clean narration is preferable to pretending a defective talking shot is accurate.
Keep the agreed wording when wording matters. Do not silently change a vow to fit the length of generated mouth movement. Secure appropriate permission for any voice material and do not assume a person's photograph authorizes cloning their voice.
Does a voice change always mean the wrong person spoke?
No. Voice identity, pronunciation and turn ownership are different checks. A consistent voice can still be attached to the wrong visible person.
Can I hide an error with captions?
Captions can clarify intended speakers, but they do not fix contradictory visible speech. If the contradiction is distracting, revise the shot or change the treatment transparently.