How to Write Scripts That Sound Natural in an AI Voice
Ten practical rules for writing text to speech scripts: short sentences, punctuation as direction, numbers spelled out and names respelled, with examples.
Short answerAn AI voice reads exactly what is on the page, so write for the ear: use short sentences, put one idea in each, and use punctuation to place pauses. Spell out numbers, symbols and abbreviations, respell names the way they sound, and read the script aloud before you generate it. If you run out of breath, the sentence is too long.
A human voice actor reads your script, understands it, and then performs it. An AI voice performs it without the middle step. Everything it knows about how a line should sound comes from the words and punctuation you give it.
That is good news, because it means delivery is something you can fix at the keyboard. These ten rules cover almost every problem.
1. Write the way you talk
Text written for the eye is denser than speech. Before you generate anything, loosen it.
- Use contractions: it’s, you’ll, we’re, don’t.
- Choose the plain word: use, not utilise; about, not approximately.
- Address one person: you, not users or customers.
Before: “Customers are advised that the application facilitates the conversion of text into audio.”
After: “You paste in your text, and the app turns it into audio.”
2. Keep sentences short
A listener cannot look back up the line. Aim for sentences of 8 to 16 words, and let a few be much shorter. A very short sentence after two longer ones sounds confident. It also gives the voice somewhere to land.
Before: “Our new blender, which has been redesigned from the ground up with a more powerful motor and quieter operation, is available from Friday in three colours.”
After: “Our new blender is here. It has a stronger motor. It is much quieter. And it comes in three colours, from Friday.”
3. One idea per sentence
If a sentence contains and, which or while more than once, it is probably two sentences. Split it. Each full stop also gives a video editor a clean place to cut.
4. Use punctuation as stage direction
Punctuation is the only direction the voice gets.
| Mark | What it does to the read |
|---|---|
| Comma | A short breath |
| Full stop | A full pause, and the pitch falls |
| Question mark | Rising intonation |
| Exclamation mark | More energy; use sparingly |
| Dash or ellipsis | A beat of hesitation or suspense |
| New paragraph | A longer pause; a change of thought |
If a line sounds rushed, add a comma. If it sounds choppy, take one out. Changing punctuation and regenerating is the fastest fix there is.
5. Spell out numbers, symbols and units
Do not make the voice guess.
| Instead of | Write |
|---|---|
| 2026 | twenty twenty-six |
| £19.99 | nineteen pounds ninety-nine |
| 50% | fifty per cent |
| 3–5 days | three to five days |
| 10kg | ten kilograms |
| & | and |
6. Expand abbreviations
Write for example instead of e.g., that is instead of i.e., and Doctor instead of Dr where it matters. For initialisms you want read letter by letter, add spaces or full stops: F. A. Q. For ones you want read as a word, such as NASA, leave them as they are and check the result.
7. Respell names and unusual words
The audience hears the audio, not your spelling, so spell for sound.
- Siobhan becomes Shiv-awn
- Nguyen becomes Nwin
- Worcestershire becomes Wuss-ter-sher
Generate the sentence, listen, and adjust. Keep a list of the spellings that work, so every script in a series pronounces the name the same way.
8. Put the important word last
In spoken English, emphasis falls naturally at the end of a sentence. Arrange the sentence so the word that matters is there.
Weaker: “A free gift comes with every order this weekend.”
Stronger: “This weekend, every order comes with a free gift.”
9. Break long scripts at natural joins
Voice Labs takes up to 5,000 characters in one generation. If your script is longer, do not cut it at the limit. Cut it where a listener would expect a pause: the end of a scene, a section or a topic. Use the same voice and settings for every part, and the joins will not be audible.
10. Read it aloud once before you generate
This is the oldest rule in broadcasting and it still works. Read the script out loud at speaking pace. Wherever you stumble, run out of breath or lose the thread, the voice will have the same trouble. Fix those places first.
A before-and-after
Before (written for a web page):
Introducing the Aurora 2 (available in 3 colours from 14/11) – our lightest, quietest & most powerful blender yet, w/ a 1,200W motor and a 5-yr warranty.
After (written for a voice):
Meet the Aurora Two. It is our lightest blender yet. Our quietest. And our most powerful, with a twelve-hundred watt motor. It comes in three colours, with a five-year warranty. Available from the fourteenth of November.
Same facts. The second version gives the voice short sentences to land on, numbers it cannot misread, and the date as the last thing the listener hears.
Then let the settings do the rest
Once the script reads well, use the voice settings for the final ten per cent. In Voice Labs, the Narration preset keeps a long read steady, Expressive adds range to short, energetic pieces, and the Speed slider handles pace. The guide to stability, similarity and speed explains what each one changes.
