Drafting Voiceover Copy That Matches How the Creator Actually Talks

I direct voiceover for a living, and the fastest way to spot AI-drafted copy is to hear someone read it aloud. The sentences are balanced. The clauses are parallel. Every thought resolves neatly. And it sounds nothing like how the person in front of the mic actually talks, because real speech is lopsided, front-loaded, and full of sentences that change their mind halfway through.

The problem isn't that AI writes badly. It writes clean written English. Voiceover isn't written English read aloud — it's a transcript of thinking out loud, shaped so it lands. If you want a draft that a creator can perform without fighting it, you have to work against the model's instinct toward tidy prose. Here's how.

Start from how they already sound

Every creator has a rhythm, and it's sitting in their existing videos. Before writing a word of new copy, I pull transcripts of two or three pieces where the person sounds most like themselves — usually unscripted talking-head segments, not their polished intros. Then I read for the patterns:

• Sentence length and variance. Do they run long and then snap short? Are they mostly clipped? The pattern of long-then-short is a fingerprint.
• Filler and connective tissue. "So," "look," "here's the thing," "right?" — the little moves they use to start and steer a thought.
• Where the emphasis lands. Do they front-load the point and then explain, or build to it?
• Contractions and register. "Do not" versus "don't" is the difference between a spokesperson and a person.

That read gives me the target. Now the model has something to match instead of defaulting to essay prose.

Give the model the voice, not just the topic

The single biggest quality jump comes from feeding real transcript samples into the draft prompt as the style reference. Not a description of the voice — the actual voice.

Here are three transcripts of how this creator talks
when unscripted. Study the rhythm: sentence length,
filler words, where they put emphasis, contractions.

[paste 3 transcript excerpts]

Now write a 45-second voiceover on [topic] that sounds
like the SAME person talking. Rules:
- Vary sentence length hard. Some very short. A couple long.
- Write for the ear: contractions, everyday words.
- Front-load the point, then explain it.
- It's fine to start a sentence with And, But, or So.
- No parallel-structure lists. No neat three-part phrases.
- Read it aloud in your head; if a phrase is a tongue-
  twister or needs a breath mid-clause, rewrite it.

The "no neat three-part phrases" rule matters more than it looks. Models love the rhythm of three — "faster, cleaner, smarter." People use it occasionally; models use it constantly, and it's the tell that gives away a script. Ban it and you force more natural, uneven phrasing.

Write for breath

Voiceover is governed by lungs. A sentence that looks fine on the page but has no natural place to breathe will trip the reader every take. When I draft or edit VO copy, I mark where the breaths go, and if a clause runs longer than a comfortable breath, it gets broken up — even if that means a grammatically "incomplete" line.

This is why VO scripts should be formatted in speech units, not paragraphs. One thought per line. Line breaks are breath cues. It looks strange as text and reads perfectly at the mic:

Okay, so here's what nobody tells you.
You don't need the expensive one.
You really don't.
The cheap version does the same job —
it just doesn't look as good on a shelf.

That's the same information a model would pack into two smooth sentences. Broken into breath units it becomes performable, and the short "You really don't" is doing emotional work that a subordinate clause never could.

Read every draft out loud — that's the test

There is no substitute for this and no tool that replaces it. Read the copy aloud at performance pace. The tongue-twisters surface immediately: clusters of hard consonants, three prepositional phrases in a row, a word that's fine on paper and awkward in the mouth. Sibilance stacks — too many s-sounds close together — hiss on a real mic. You catch all of it by reading, none of it by looking.

If you're drafting for someone else to perform, read it in their cadence, not yours. And when you can, have the creator read a draft and mark anything they'd never say. That list becomes a permanent style note — feed it back into future prompts as "this person never says: X, Y, Z." The voice sharpens every cycle.

Where AI voices fit, and where they don't

The 2026 synthesis tools — ElevenLabs and the rest — will read your copy in a cloned or stock voice convincingly enough for a lot of uses. Two honest cautions, and I'll keep them brief since they sit outside pure scriptwriting.

First, the copy still has to be written for speech. A synthetic voice reading essay prose sounds exactly as wrong as a human reading it — the model doesn't fix bad rhythm, it performs it faithfully. Everything above applies whether a person or a synthesizer is reading.

Second, and this is a genuine caution rather than legal advice: cloning a voice is only yours to do with clear consent from the person whose voice it is, and the platform terms and disclosure norms around synthetic voice are still shifting in 2026. If you're cloning your own voice for your own channel, that's straightforward. If it's anyone else's, get explicit permission in writing and check the current platform rules before you publish. I'm flagging the practice, not giving you a compliance ruling — treat it as a prompt to check, not a green light.

The draft-to-performable loop

1. Pull transcripts of the creator sounding most like themselves.
2. Feed those samples as the style reference, not a description.
3. Draft with hard rules against tidy prose — length variance, no rule-of-three, sentence-starting conjunctions allowed.
4. Reformat into breath units, one thought per line.
5. Read it aloud at pace; kill tongue-twisters and sibilance stacks.
6. Have the creator flag anything they'd never say, and bank those notes for next time.

The model gets you 80 percent of the way in seconds, and that last 20 percent — the breath, the lopsidedness, the small words that make it sound like a person — is the part that decides whether anyone believes it. That part is still yours.