Before you generate AI narration, run twelve checks on the script. Read it aloud, shorten long sentences, write numbers and acronyms as spoken, put names in a pronunciation dictionary, mark speakers only where they change, use pause pills for exact silence, keep notes out of the spoken text, and generate one test passage first.
Most problems you'd hear in a finished read are cheaper to fix in the text than in the edit. The full background for each check is in our primer, How to write a script for AI voices.
The 12-point checklist
| # | Check | Why |
|---|---|---|
| 1 | Read it aloud | Stumbles on the page become stumbles in the audio |
| 2 | One idea per sentence | Listeners can't scroll back |
| 3 | Vary sentence length | All short sounds clipped; all long is hard to follow |
| 4 | Numbers, dates and units as spoken | "03/04" can be read more than one way |
| 5 | Acronyms decided | Letters (A P I) or a word (NASA) |
| 6 | Names in the pronunciation dictionary | Fixed once, everywhere |
| 7 | Speakers marked only where they change | Fewer pills, fewer mistakes |
| 8 | Exact silences as pause pills | Punctuation doesn't set a length |
| 9 | Headings and notes not in the spoken text | The voice reads what it's given |
| 10 | Reaction tags only where the engine performs them | Unsupported tags do nothing |
| 11 | Links, paths and symbols rewritten | "example dot com slash help" |
| 12 | One test passage generated and measured | Catch problems before the full run |
1. Did you read it aloud?
Out loud, not in your head. Mark every place you trip, run out of breath or lose the thread, and rewrite it. If you stumble, the voice probably will too.
2 and 3. Is each sentence one idea, with varied length?
Split sentences that carry two or three ideas. Then check the rhythm: a run of very short sentences sounds clipped, and a run of long ones is hard to follow. Mix them. Swap written transitions such as "Consequently" for spoken ones such as "So" or "And that's when". Use contractions where the tone is conversational.
4 and 5. Are numbers, dates and acronyms written as spoken?
Write what you want said: "eleven at night", not "11pm"; "April third, twenty twenty-six", not "03/04/2026"; "twelve dollars and fifty cents", not "$12.50". Decide for each acronym whether it's spelled out ("A P I") or said as a word ("NASA"). Full guide: How to write numbers, dates and acronyms for AI voices.
6. Are the names in your pronunciation dictionary?
For people, places, products and brand terms that repeat, add a rule in Settings > Pronunciation instead of respelling the word in every script. Choose This project only or Every project, and use Import CSV to load a whole glossary. See the pronunciation dictionary guide.
7. Are speakers marked only where they change?
In Vois Studio, type / and choose Speaker to add a speaker pill. Every line after it belongs to that speaker until the next pill, so you don't need to repeat it on each line. Check that every change of voice has a pill and that no line has a wrong one.
8. Are exact silences pause pills?
Punctuation shapes rhythm, but it doesn't set a length. Where a silence matters, such as before a reveal, between sections or while the screen changes, type /, choose Pause and pick a length from 0.3 to 3 seconds. Markup <break> tags aren't supported. More: Pause nodes for precise silence.
9. Is everything that shouldn't be spoken out of the spoken text?
A voice reads what's in the line. Move chapter names, timing notes, "say this warmly" reminders and on-screen cues into Heading, Title, Section or Note blocks from the / menu. They stay visible in the editor and are never spoken.
10. Are reaction tags only where the engine performs them?
Reaction tags are sounds, not moods. In Vois Studio, Expressive performs nine English tags, such as [laugh] and [sigh], where you place them. Fast and Multilingual remove tags before speaking, so you hear neither the sound nor the brackets. There's no [happy] or [sad]; put feelings in the words. Use tags sparingly, one sound per spot.
11. Are links, file paths and symbols rewritten?
URLs, file paths, code, hashtags and symbols are written for eyes. Rewrite them as you'd say them, or move them to the video description or show notes and say "link below".
12. Did you generate and measure a test passage?
Pick a passage that has the hardest parts: names, numbers, a speaker change, a pause. Generate it in the voice you'll use, listen with the script in front of you, and fix what you hear. Measure its length too, and plan the full script from that rate: How long is a script read aloud?
Related
- How to write a script for AI voices
- How long is a script read aloud?
- Pronunciation dictionary guide
- How to write numbers, dates and acronyms for AI voices
- Pause nodes for precise silence
Print the table, tape it next to your screen, and tick all twelve before the next long script goes in.
The Vois Team