Fixing one word in a sung demo with AI: the free way first, the paid tools second, and the consent line

To change one word in a sung demo, the free answer is usually the best one: sing the new word, punch it in and comp it so the join is invisible. AI helps only when you cannot re-sing it yourself, and then the job is voice conversion (making a new take sound like the original singer), which you should do only with a voice you own or have permission to use. Everything here is as we read the sources on 29 Oct 2026.

Free first: re-sing and comp

One word is a two-minute fix in any DAW, including free ones. Record the line again with the same mic, distance and room, cut the new word in at a consonant, and crossfade. Our guide to vocal editing, comping and timing covers where to cut and how to hide the join. If the tone does not match, a little EQ usually closes the gap; see vocal chain basics.

What we think: most "AI inpainting" searches are really this problem. A real take of your own voice still sounds more like you than any model does.

When you cannot re-sing it

Typical cases: the singer is unavailable, the session is gone, or you only have a bounced mix. The AI workflow then has three steps:

1. Isolate the vocal. If you only have the mix, separate it first. The free RVC project includes vocal/accompaniment separation, and our stem separation guide explains the artefacts to listen for.
2. Sing a guide for the new word. Anyone can sing it; the phrasing and pitch come from this take.
3. Convert the guide into the original voice, then comp it in as above.

The free AI option: RVC

RVC (Retrieval-based Voice Conversion) is free and MIT-licensed. The project says you can train a good model on 10 minutes or less of voice data, and recommends collecting at least 10 minutes of low-noise audio. It also ships a real-time voice-changing interface. The cost is setup time: you install it locally and need a capable computer. How it works is in AI voice conversion explained.

For one word, you train on the singer's clean takes from the same project, convert your guide word, and drop it in. Expect to try a few renders and edit the breaths by hand.

Paid tools

Commercial vocal tools wrap the same idea in an app. ACE Studio's pricing page lists paid "rent-to-own" plans that include voice-cloning slots for both its vocal synth and its vocal changer, with more slots on the higher tier. Prices were shown with a holiday discount when we looked, so check the live page before buying. Pay only if RVC's setup is the real blocker.

The consent line

The technique that fixes your own demo can put words in someone else's mouth. Our rules:

• Your own voice: go ahead.
• A collaborator or session singer: ask first, in writing, and say it is a replaced word, not a re-sung one. See how to license your voice for AI for what terms to agree.
• Anyone else: don't. Training a model on a stranger's stems is the cloning problem covered in how to protect your voice from AI cloning.

Disclosure matters too. In the EU, Article 50 of the AI Act requires deployers of AI-made or manipulated audio that counts as a deep fake to disclose that it is "artificially generated or manipulated"; for evidently artistic or creative work, the duty is limited to disclosing that such content exists, in a way that does not spoil the work. Whether one swapped word in your own voice meets that bar is a legal question we cannot settle, but if you release the track, our guide to disclosing AI use through your distributor shows where to say so.

What we think: one converted word in your own voice is editing, much like tuning. One converted word in someone else's voice is a decision they should get to make.