Writing a script for an AI voice: punctuation, numbers, pace
12 October 2026 · 6 min read
An AI voice reads exactly what you write. That is its strength and its trap: it does not guess your intentions, it reads them in the punctuation, the word order and the way you write numbers. Here are the rules that change everything, tested on Aievalu's English and French voices and valid on most AI voice generators.
1. One idea per sentence
Long sentences with three subordinate clauses read fine on paper and badly to the ear. Cut. Twelve to twenty words, one verb, one idea. The listener cannot scroll back: if a sentence needs re-reading, it will be misheard.
Before: "In this video we are going to look at how to set up your account, which is fairly simple if you follow the steps I am about to describe, before moving on to the advanced options."
After: "In this video, we set up your account. It's simple. Then, the advanced options."
2. Punctuation is your mixing desk
A full stop is a clean pause and a falling intonation. A comma is a short pause. An ellipsis slows things down. A question mark lifts the end of the sentence. A blank line between paragraphs creates a longer breath. If a sentence is read too fast, add a comma; if two ideas collide, add a full stop. Do not use capitals to shout: the voice cannot hear them, and some engines spell them out.
3. Numbers: write what you want to hear
Recent engines read numbers well, but each form has its own reading. "1500" becomes "fifteen hundred" or "one thousand five hundred" depending on the voice; "$1,500" becomes "one thousand five hundred dollars". For a date, "12 October 2026" is safe; "10/12/2026" is ambiguous between US and UK readings. Write a phone number in groups, "020 7946 0018", to get the natural rhythm. Write "15%" for a percentage, "3.5" for a decimal. When the reading matters, spell the number out in words: it is the only unambiguous option.
4. Acronyms and foreign words
"NASA" is pronounced as a word, "HTML" is spelled out. If the engine gets it wrong, force it: "H.T.M.L." to spell, "Nasa" for a word. For a brand name, test both spellings and keep the one that sounds right; that is what we did for our own name. French words in an English text are read with the voice's accent: if "rendez-vous" sounds off, accept it or rephrase. An English voice will never read a whole French sentence correctly: use the voice of the text's language, our English and French voices exist for that.
5. Dialogue and tone
One voice does not play several characters. For a dialogue, generate each line with a different voice, or a different tone of the same voice, then assemble in your editor. Tone is chosen before generating: Paul "Cheerful" will not read a maintenance notice like Paul "Neutral". Emphasis comes from word order, not bold type: "It's free, and no account needed" stresses free; "No account needed, and it's free" stresses the absence of an account.
6. The 1,500-character limit, and how to cut
One Aievalu generation accepts up to 1,500 characters, about a minute and a half. That is not an arbitrary constraint: it is the right unit of work. Cut your script into blocks of one idea or one scene, one block per paragraph. Each block becomes a file you can regenerate on its own if an intonation is wrong, without touching the rest. Name the files in order (01-intro, 02-context…), keep the same voice and tone from one block to the next, and assemble in your editing software.
7. Listen, fix, regenerate
The first take is rarely the right one, and that is fine: credits are only charged for what is produced, and regenerating a 300-character passage costs under 3 cents. Listen on headphones, note the three places that snag, fix the text (a comma, a number in words, a moved word) and rerun that block alone. Three iterations are almost always enough. The usage rights on the result are detailed in the FAQ.
8. A checklist before you generate
- Short sentences, one verb, one idea.
- Deliberate punctuation: full stop = pause, comma = breath.
- Numbers written the way you want them heard.
- Acronyms and brand names tested, spelling frozen.
- One block per idea, 1,500 characters max, same voice throughout.
- Text read aloud once: if you stumble, the voice will stumble.
To see these rules applied to a specific format, read our YouTube, podcast and e-learning guides, or test two sentences of your script right now with the free trial. Pricing starts at €9 for about a hundred minutes.