MiniMax H3 dialogue replacement for AI video scenes

MiniMax H3 dialogue replacement combines a visual edit with a performance and sound task. A new line needs the correct speaker, timing, emotion, lip movement, room acoustics, and relationship to the existing ambience. The launch manual supports audio references in Omni Reference and describes dialogue-oriented editing examples, giving creators a way to specify vocal tone alongside the source video.

Treat the original clip as the timing and environment record. Decide which parts of its soundtrack should change before writing the new line.

Supplied H3 demonstration: a new referenced line is paired with adjusted visible performance.

Prepare the line and voice role

Write the exact dialogue, including names and product terms. Add direction for pace, emotional shift, volume, pauses, and who the speaker addresses. Keep the line realistic for a 5-to-15-second result.

An MP3 or WAV reference can define vocal quality or delivery. Each audio reference must be 2 to 15 seconds, and audio references total at most 15 seconds in a request. Audio cannot stand alone, so pair it with the source video or an image.

If the scene has multiple speakers, make the order explicit. Identify whether the other voices remain unchanged or whether silence replaces an original response.

Protect the sound bed

List the audio that should survive: room tone, traffic, music, footsteps, or object sounds. When the old dialogue overlaps these elements, describe the desired final mix rather than assuming the model will know which layer to retain.

Also protect the visible performance where possible. If the line changes length substantially, specify whether timing, facial expression, or the edit may adapt.

Review in three passes

First listen without watching for words, pronunciation, speaker identity, emotional delivery, noise, and stereo balance. Then watch silently for mouth shapes, facial effort, gaze, and gesture. Finally combine picture and sound to judge sync and whether the voice seems located in the room.

Check the beginning and end for clipped syllables. For a series, record approved pronunciations and reuse consistent reference material.

Frequently asked questions

What audio formats does MiniMax H3 accept?

The manual lists WAV and MP3 for standalone audio references.

Can an audio reference replace the source video?

No. Audio cannot be the only reference input; it must accompany an image or video.

Does a new line automatically preserve background sound?