Edit an AI influencer video with an optional voice

Disclosure: some links in this tutorial are affiliate links. I may receive a commission if you subscribe through them. The examples and limitations come from our documented experiments.

A voice is optional. We finish the files with Codex: keep or restore source sound, join clips when needed and check the result. The following sections help if you want to add a spoken line or change a voice.

Listen to the soundtrack without watching first, then watch with sound to check the cuts. If there is a spoken line, make sure it is complete and audible over the music. Check the lips only when the character speaks on camera.

For this stage, keep the clip you accepted after the Genjutsu replacement.

The method: approve an output at every stage

Prepare the final file with Codex

In our tests, Codex helped restore source audio and assemble files. We reapplied the original audio to Glen’s tiny car clip and checked its duration. For the kiss scene, we joined a five-second extension and its music. Describe your files and the result you want to request the same operations.

1. Put the source clip and accepted result in a folder Codex can access. Give them distinct names.
2. Ask it to compare durations and audio tracks. If the source sound works, ask to restore it to the result and explain any offset or cut needed.
3. Watch the entire resulting MP4. Check the start, end and every join. Keep the source files for corrections.

Adapt this instruction to your filenames. It describes the result and lets Codex choose the local operations needed.

Compare source.mp4 and result.mp4 in my folder. Report their duration, dimensions and audio tracks. Make final.mp4 from the accepted result, preserving its framing. Restore source.mp4 audio if the result did not keep it correctly. If durations differ, explain the required cut or join before applying it, without stretching the voice. Add no new voice or text. Check the start, end, audio synchronisation and any joins, and keep the sources.

If the sound works, you can stop here. ElevenLabs is optional for adding a voice or changing a spoken line. The following sections cover those cases.

Choose the right voice operation

Text-to-speech reads a script. Voice changing transforms an existing recording. Lip synchronisation adjusts visible mouth movements. ElevenLabs provides text-to-speech and Voice Changer. Replacing an audio track does not automatically synchronise the whole video.

1. Write a short line that fits the character.
2. Choose a suitable voice and test one sentence.
3. Check names, pronunciation, pauses and the final word.
4. If the mouth is visible, prepare a separate lip-sync step and review its result.

What Gérard taught us

Gérard’s parody about AI influencers | YouTube

French dialogue, original video.

We explored Jayce, Jean Maurice and Zion Aquila, with several timing and lip-sync versions. The first public parody combines a Genjutsu replacement, generated speech and synchronisation. Those stages must be attributed separately.

Gérard’s later original interview uses Seedance 2.5 with a character reference and written dialogue. A local VoiceStudio recording served as a reference, while the model’s native speech was retained. It is a different workflow from adding the first parody’s voice afterwards.

Finish the edit

Keep useful ambience, lower music under speech and listen to each cut. Glen’s car scene had text removed with FLUX and was then extended for the kiss. The retouched version changed resolution and duration. The boxing extension kept native sound. The requested shout was approximate. Neither example is the work of one swap operation alone.

Check the exported file, not only the editor: the start, middle and end. Check the face and, if the character speaks on camera, the lips. Sound levels. Unwanted text. Duration. Vertical framing. Naïla’s restored audio extended 0.168 seconds beyond the final video frame, a small discrepancy worth checking.

Return to the complete tutorial or revisit the Genjutsu swap checks.