AI Voice Inpainting: How to Change Words in Speech Without Re-Recording
Last updated 2026-10-14
AI voice inpainting is the audio version of image inpainting: you mark a region — say, one wrong word — and a model fills it with new speech that matches the voice, pace and room tone around it. The rest of the recording stays exactly as it was.
The terms people use
- Voice / speech inpainting — the research and API term.
- Word-level editing — editing audio by editing words in a transcript.
- Overdub / regenerate — product names for typing new words in a cloned voice.
- “Change one word in audio” — what most people actually search for.
How it works
1. Transcribe with word timestamps
Speech recognition produces the transcript plus a start and end time for every word, so a click on a word maps to an exact slice of audio.
2. Pick the region to regenerate
Regenerating a lone word usually sounds unnatural — intonation depends on the whole sentence. Good tools expand the edit to the surrounding phrase or sentence.
3. Generate in the same voice
The new text is synthesized conditioned on the original speaker — either true inpainting (conditioned on audio on both sides of the gap) or zero-shot voice cloning from nearby audio.
4. Blend it back
The new audio is loudness-matched, trimmed and crossfaded into the original at natural pauses so there's no click or jump in level.
Inpainting vs. regenerating the whole clip
| Regenerate everything | Inpaint one region | |
|---|---|---|
| Untouched parts | Delivery changes everywhere | Identical to the original |
| Cost | Pay for the whole script again | Pay for one sentence |
| Works on real recordings | No | Yes |
Try it
WordRetake does this in the browser: upload a file, click the word, type the fix, download. Comparing tools? See 5 AI voice tools that can change a single word.
Fix a word in your own audio — free
Upload, click the word, type the fix. Two edits free, no sign-up.
Open the editor