# How to write a voiceover script for an AI voice

Updated 2026-11-06

Write a voiceover script for an AI voice the way you'd say it out loud. Start with the word count, sized to the voice's pace. My voices measured about 178 words a minute on average, so a sixty-second spot on them gets roughly 175 words. That is more than the 150 often quoted for voiceover. Then keep one thing per sentence, spell every number the way it should sound, and build the pauses out of punctuation. Get those four right and most of what people call "robotic" is gone before you press generate. Below is how I write scripts for my own voices, with a paste-ready example near the end.

## How many words should a voiceover script be?

Pick the length first. The word budget follows from it. These come from timing my own Eleven v4 clips, grouped by pace, and [I measured my voices at 178 words per minute](https://voicefieldguide.com/research/elevenlabs-words-per-minute) overall.

| Pace | Words per minute | 30 seconds | 60 seconds | 3 minutes |
|---|---|---|---|---|
| Quick (Pulse, Reel, Dexter, Timber, Bilbop) | about 190 | 95 | 190 | 570 |
| Medium (Snap, Derek, Jett, Heisenberg, Havoc, Reverend, Felix, Titan) | about 175 | 87 | 175 | 525 |
| Unhurried (Titan Epic, Drift) | about 155 | 77 | 155 | 465 |

Tutorials and screen walkthroughs need about 20 percent fewer words, so the viewer gets quiet beats to watch each click. All of it is a starting point. The first generated take tells you the real timing, and the [words per minute calculator](https://voicefieldguide.com/tools/words-per-minute-calculator) runs the arithmetic in either direction.

## How do I write a voiceover script that sounds spoken?

Say every line aloud as if you were talking to one person across a table. If a sentence needs a second breath, it's too long for the listener as well.

- **Vary sentence length.** A short line after a long one gives the voice a place to land. Six medium sentences in a row make the metronome rhythm people blame on the AI.
- **Use contractions.** It's, you'll, don't. Drop them when the register turns formal, as in a bulletin or a sermon.
- **One idea per sentence in instructions.** Someone following along has their hands busy and cannot scroll back.
- **Reach for concrete nouns.** "The blue Export button" gives a voice something to read. "An intuitive interface" gives it nothing to hold.
- **Cut the stock openers.** Anything you'd skip on a podcast, the listener skips too. An AI voice reads filler with complete sincerity, and that somehow makes it worse.

## How should numbers be written in a voiceover script?

Write them as words. Numbers, dates, currencies and acronyms are where AI reads most often go wrong, and ElevenLabs' own docs show it happening. One of its models reads "$1,000,000" as one million dollars, while a smaller model reads it as one thousand thousand dollars. Don't leave that to chance.

- 2026 becomes "twenty twenty-six"
- 3.5% becomes "three point five percent"
- SKU becomes "S-K-U" if you say the letters
- Dr. becomes "Doctor"

When a name keeps coming out wrong, respell it phonetically in the script. On Eleven v4, ElevenLabs also documents IPA wrapped in forward slashes for exact pronunciation.

## How do I add pauses without break tags?

With punctuation. Eleven v3 and v4 don't support break tags. The `<break time="1s" />` tag works on other ElevenLabs models, for pauses up to three seconds, and does nothing useful on v3 or v4. On v4 you have five tools.

- **Full stops** give a clean beat.
- **Paragraph breaks** give a longer one.
- **Ellipses** slow a line or let it trail off.
- **Dashes** cut a thought short.
- **Capitals** on a single word add emphasis. Use them once or twice per script, or they stop meaning anything.

Audio tags in square brackets, like `[sighs]` or `[whispers]`, work on v3 and v4 too. Go easy on them. My [audio tags guide](https://voicefieldguide.com/guides/elevenlabs-audio-tags) covers which ones earn their place.

## What should I paste into ElevenLabs?

Only the words that should be spoken. A traditional voiceover script runs in two columns, visuals on the left and audio on the right, and it's a fine planning document. Keep it for yourself. Scene notes, speaker labels and timecodes all get read out loud if they reach ElevenLabs. On v3 and v4, anything in square brackets is treated as a direction, which means a leftover `[client name]` placeholder will not be skipped politely. Fill in every bracket before you generate.

ElevenLabs also says v4 works best on paragraphs. Generate a whole section in one go and trim it in the edit. Feeding the model one sentence at a time gives it too little to work with.

## What does a finished one-minute voiceover script look like?

Here's one sized for a one-minute slot. It runs 143 words, which comes to about 49 seconds at a medium 175 words a minute, leaving room for the pauses the line breaks create.

> Look up on a clear afternoon. The sunlight pouring down looks white, but the sky around it is another color. Blue.
>
> Sunlight is every color mixed together. When it reaches the air, it runs into gas molecules far smaller than a wavelength of light. Those molecules scatter short wavelengths much more strongly than long ones. Blue light is short. Red light is long.
>
> So blue gets bounced all over the sky... and some of it heads toward your eyes from every direction. That's the glow you see overhead.
>
> Violet is shorter still, so why isn't the sky violet? Two reasons. Sunlight carries less violet to begin with, and your eyes are far less sensitive to it.
>
> At sunset, the light crosses much more air to reach you. Most of the blue is scattered away before it arrives.
>
> What's left is orange. And red.

Each choice in it is deliberate. "Blue." gets a sentence to itself, so the voice comes to a hard stop on the payoff word. The ellipsis in the third paragraph gives the turn a breath. Paragraph four asks the question the viewer is already asking, which keeps it from sounding like a trick. And the last line is split in two because a list of colors read in a single breath goes flat.

## Why generate the same voiceover script twice?

Read the script aloud once before you spend credits. Then generate it twice. Every take is a slightly different performance, which is what the Stability setting governs, and two takes side by side show you which problems belong to the script and which belong to the take. When both stumble in the same spot, the sentence is at fault. Fix it there.

## Which voices show a script's weak spots?

**Derek** is the voice I'd use to proof a script heavy on names, dates and figures. Its firm broadcast pacing puts every number in plain view, so a badly written date has nowhere to hide. It's formal by default, though, and a chatty creator script can come out sounding like a press release. Hear it on [Derek for news and announcements](https://voicefieldguide.com/elevenlabs-voices/derek-captivating-news-anchor/news-and-announcements).

**Havoc** is my test for story scripts. Its Southern drawl and grit need stakes on the page. When the writing has them, Havoc sounds like a real place and a real person. When it doesn't, you hear the gap within a line or two. Neutral or corporate copy is the wrong home for it, because the accent pulls focus. See [Havoc for storytelling](https://voicefieldguide.com/elevenlabs-voices/havoc-gritty-southern-narrator/storytelling).

Compare [Derek](https://voicefieldguide.com/elevenlabs-voices/derek-captivating-news-anchor) on factual copy, [Havoc](https://voicefieldguide.com/elevenlabs-voices/havoc-gritty-southern-narrator) on a story opening, and [Heisenberg](https://voicefieldguide.com/elevenlabs-voices/heisenberg-polished-and-clear) on an explanation. Each sample gives the voice something from its catalog strengths to work with.

## Audition details

These picks compare my fifteen voices. I own them, and ElevenLabs may pay me when you use them.

### Derek: Factual copy

Broadcast delivery puts names and figures in view. Casual banter may sound too formal.

- Voice ID: `YgrS0YQVeTTwGJKRYx5E`
- Voice page: https://voicefieldguide.com/elevenlabs-voices/derek-captivating-news-anchor
- Add to My Voices: https://elevenlabs.io/app/voice-lab/share/986dea1fcfbe6dcbda4c7f6980fdf8f0b59fda73adb04da7dfdd4be27beac5a2/YgrS0YQVeTTwGJKRYx5E
- This guide's Eleven v4 clip: https://voicefieldguide.com/audio/v4/guide--how-to-write-a-voiceover-script--derek-captivating-news-anchor.mp3

Sample line: The first train leaves at six forty-five. The second leaves twenty minutes later, from a different platform. Check the destination on the board before you board; the two services follow the same route for only three stops.

### Havoc: Story copy

Southern grit suits folklore with stakes on the page. Neutral corporate text will pull against the accent.

- Voice ID: `dtVZnErhiiosqofxDzSH`
- Voice page: https://voicefieldguide.com/elevenlabs-voices/havoc-gritty-southern-narrator
- Add to My Voices: https://elevenlabs.io/app/voice-lab/share/1af53105b4677224ad4103bd4a69bb4519fdc844636df8bda01ed265f5c1b3d2/dtVZnErhiiosqofxDzSH
- This guide's Eleven v4 clip: https://voicefieldguide.com/audio/v4/guide--how-to-write-a-voiceover-script--havoc-gritty-southern-narrator.mp3

Sample line: That old gate had been chained shut since I was a boy. I knew the sound it made in the wind. Three short knocks, then a scrape. So when I heard four knocks that evening, I went to the window before I reached for the light.

### Heisenberg: Dense explanations

Even educational delivery gives a complicated sentence space. It reads near the middle of my range, about 175 words a minute.

- Voice ID: `iEBOK9alpKauGRvBSsFi`
- Voice page: https://voicefieldguide.com/elevenlabs-voices/heisenberg-polished-and-clear
- Add to My Voices: https://elevenlabs.io/app/voice-lab/share/e0464333b21941625fcdd28b2ce050bc74911f6bd69cec0d6861a8bd9e24032b/iEBOK9alpKauGRvBSsFi
- This guide's Eleven v4 clip: https://voicefieldguide.com/audio/v4/guide--how-to-write-a-voiceover-script--heisenberg-polished-and-clear.mp3

Sample line: Read the sentence once without the phrase in brackets. If the meaning survives, leave that phrase out of the spoken version. The listener gets one pass at the explanation, and every extra clause asks them to hold another thought.

## Sources

- [ElevenLabs docs: Best practices (pauses, text normalization, pronunciation)](https://elevenlabs.io/docs/overview/capabilities/text-to-speech/best-practices)
- [ElevenLabs docs: Eleven v4](https://elevenlabs.io/docs/overview/capabilities/text-to-speech/eleven-v4)
- [ElevenLabs Help Center: How do audio tags work with Eleven v3 and v4?](https://elevenlabs.io/docs/help-center/product/core-capabilities/text-to-speech/how-do-audio-tags-work-with-eleven-v3-and-v4)
- [ElevenLabs: Emotional text to speech with Eleven v4](https://elevenlabs.io/blog/emotional-text-to-speech-with-eleven-v4)

Source page: https://voicefieldguide.com/guides/how-to-write-a-voiceover-script
