How to write a script for AI voiceover

The format abcVoiceover reads

A script is a list of lines, and each line is one take and, in the download, one audio file. When you paste a script, start a line with the name of a voice in the project's cast and a colon:

Narrator: The station had been silent for three days.
Mara: Hello? Is anyone there?
Joss: Mara. You shouldn't have come back.

Each of those lines goes to the voice with that name. A line without a name goes to the voice picked beside the box you paste into, and you can change any line's voice afterwards. A line can be up to 2,000 characters.

How long should a line be?

One to three sentences is the sweet spot. Long lines work, since abcVoiceover makes them a few sentences at a time and joins the pieces, but a short line is easier to direct: if one word is wrong, you re-take one sentence, not a paragraph. For narration, one line per paragraph or per slide keeps files easy to match up later.

Punctuation is direction

Voice models take their timing from punctuation, so use it the way an actor would read it:

• Full stops give a clear pause and a falling ending. Use more of them than you would on the page.
• Commas are short breaths. Add one wherever a speaker would naturally pause.
• Question marks usually lift the end of a line; make sure questions end with one.
• Exclamation marks add energy. Use them sparingly, or every line sounds shouted.
• Dashes and ellipses suggest an interruption or a trailing-off. Results vary, so generate a couple of takes and keep the best.
• Avoid ALL CAPS for emphasis; it can be read as an abbreviation. Rephrase, or split the line, to move the stress.

Spell it the way it's said

A voice model reads exactly what's written. Anything a person would have to interpret, write out:

• Numbers and dates: twelve fifty or one thousand two hundred and fifty, not 1,250; March third, not 3/3.
• Money, units and symbols: forty percent, five dollars, and instead of &.
• Acronyms: N-A-S-A or NASA depending on how it's said; F-A-Q or fack.
• Invented and unusual names: spell them phonetically in the script, like Ee-lay-nuh for Elena if it keeps coming out wrong.
• Web addresses: example dot com slash help.
• Words with two pronunciations (read, lead, live, wind): if one comes out wrong, rephrase the sentence so only one reading makes sense.

Keep everything else out of the line

Everything in a line is read aloud. Stage directions, sound effects, scene headings and notes like (angrily) belong somewhere else: in your own copy of the script, or in a separate script you never generate. To change how a line is delivered, use the voice's style instead: Expressive for more emotion, Calm for a steadier read, or Fine-tune for emotion intensity and expressiveness.

Write for the ear, not the page

• Prefer short sentences and plain words. Long, nested sentences are hard to follow when heard once.
• Use contractions (you're, it's, we'll). They sound like people talking.
• Put the important word at the end of the sentence, where the voice naturally lands.
• Read the script aloud yourself before generating. Anywhere you stumble, the voice probably will too.

Then iterate line by line

Generate the whole script, listen through, and fix only what needs it. If a word is mispronounced, change the spelling in that line and generate it again. If the delivery is off, generate another take and compare it with the current one. When a line is right, save it: it keeps that take, its text and the voice that made it, whatever changes later.

Checklist

• Every line starts with a voice name and a colon, or is meant for the voice picked beside the paste box.
• Lines are one to three sentences, with punctuation at every pause.
• Numbers, symbols, acronyms and names are written the way they're said.
• No stage directions or sound cues inside spoken lines.
• You've read it aloud once.

Next

• How to record a voice sample for AI voice cloning
• What people make with abcVoiceover