How to remove a narrator who talks over a character's line
Tell Vizard Agent which voice should go and what must stay. It splits the music from the voices, then separates the two voices from each other with a speech-separation model and keeps only the character's line. It checks the result by transcription and listening, then raises the music, which sounds thin once the narration is gone.
What do you type into Vizard Agent?
Name the voice to remove by its role, the narrator, rather than by a time, and list everything that must survive: the other voices and the music. If the overlap is the problem, say so, because it tells Vizard Agent that muting the narrator's sections would cut the character too.
Remove the narrator's voice, because the voices overlap. Keep the other voices and the music. First try to clean the audio and cancel the overlap.
What does Vizard Agent actually do?
It treats the soundtrack as three layers: music, narrator and character. Two people talking at once is the hard part, because ordinary noise removal treats both as speech and keeps both. In the session behind this article it:
- Listened to the 18-second clip and transcribed it, confirming a narrator who ran almost the whole way and a character's line underneath.
- Separated the music from the voices with a music-separation model and confirmed the music layer held no speech.
- Ran a two-speaker speech-separation model on the voice layer and checked each result by transcription: one held only the narration, the other only the character's line.
- Because that model works at a low sample rate, used its results as a guide to filter the full-quality voice, only inside the character's window, so narration elsewhere stayed out.
- Measured the line at −16.2 LUFS and the music under it at −43.5, raised the music by 6.5 dB, and told the user a slight colouring of the voice remained.
When does this not work well?
When the two voices sound alike. Speech separation tells speakers apart by the sound of their voices, so two similar voices, or heavy effects on top of both, leave traces of one in the other. Then muting the overlap or re-recording the line is the cleaner choice.
Common questions
Can two people talking at once be separated? Often, yes. Vizard Agent checks each result by transcription.
Will the music survive? Yes. Vizard Agent splits it off first and rebalances it.
Will the voice sound different? Slightly. Vizard Agent tells you when colouring remains.