How to make a character's narration and on-camera line sound like the same person
Tell Vizard Agent the character's voice changes between the narration and the line said on camera. It records the whole script, including the on-camera line, as one generation so the voice cannot drift, replaces the video generator's own voice in the talking shot, and lays the new read onto that shot's original word timings.
What do you type into Vizard Agent?
Say that every line should sound like one person, and name any line that still sounds machine-made. The mismatch is easy to miss in production because each line sounds fine on its own, and it only shows when the narration hands over to the character speaking on screen.
His voice doesn't sound like the same person. I want it softer, and every line should sound like the same person. "See you at 3" in particular sounds robotic.
What does Vizard Agent actually do?
It finds where the second voice came from. Narration recorded in several separate generations varies a little each time, and a talking shot voiced by the video model from a reference only approximates the voice, so the fix is one source for everything. In the session behind this article it:
- Found the narration had been generated as five separate pieces at very different levels, while the on-camera line came from the video generator.
- Recorded the full script, on-camera line included, as one take, rejected a voice with heavy vocal fry, and kept one judged a consistent single speaker.
- Split that take at the transcript's boundaries and levelled each piece.
- Replaced the talking shot's audio after a lip-sync pass failed review, laying the new read onto the shot's original word timings instead.
- Took the compression off the voice and re-recorded only the robotic closing line a little slower, confirming the rest of the mix was untouched.
When does this not work well?
When the on-camera line is long. A new read laid onto the old word timings holds for a short line, but over several sentences the pauses drift apart from the mouth, and a proper lip-sync or a cutaway is needed. A voice library with no close match also limits how natural the single take can be.
Common questions
Why does my character sound like two people? Two sources voiced it. Vizard Agent records every line in one take.
Can it fix just one robotic line? Yes. Vizard Agent re-records that line and leaves the rest.
What if lip-sync fails? Vizard Agent fits the new read to the shot's own word timings.