Write voiceover to be heard once, at speed, by someone doing something else. That single constraint produces every rule that follows: short sentences, one idea each, concrete nouns, and no clause a listener has to hold in memory to understand the end of. The word-count math is fixed and worth memorizing: comfortable narration runs about 140 to 160 words per minute, so a 30-second script is roughly 75 words and a 60-second script is about 150.
The rules
One idea per sentence. Written prose can carry two clauses. Spoken narration cannot, because the listener has no way to go back.
Front-load the sentence. The important word goes early. A listener who checks out halfway through still gets the point.
Concrete over abstract. "Cuts your setup from an hour to four minutes" beats "streamlines your workflow." Abstractions are invisible when heard.
Read every line aloud. The only real test. Anything you stumble over, the reader will stumble over, and a listener will hear the stumble as doubt.
Write the silences. Voiceover that runs continuously exhausts. Mark the pauses; they are where the picture does the work.
Do not narrate the picture. If the frame shows someone opening a package, the line should not say "she opens the package." Voiceover and picture should do different jobs.
The word-count math
| Runtime | Words at ~150 wpm | Realistic with pauses |
|---|---|---|
| 15s | ~38 | 30-35 |
| 30s | ~75 | 60-70 |
| 60s | ~150 | 120-140 |
| 90s | ~225 | 190-210 |
Always write to the lower number. Scripts written to the maximum produce a read that races, and racing narration is the most common reason a good script sounds cheap.
Writing for AI narration specifically
Synthetic voices have gotten good enough that the constraints are now stylistic rather than technical, with three exceptions worth writing around.
Punctuate for breath, not for grammar. A comma is a short pause and a period is a longer one, so punctuation is your timing control. Full stops where you want a beat, even when a comma would be technically correct.
Spell out anything ambiguous. Numbers, units, acronyms, and product names with unusual capitalization. Write "twenty twenty six" if you want it read that way, and "A P I" if you want the letters.
Avoid long parenthetical clauses. Synthetic reads handle them less gracefully than a human, who would use pitch to signal the aside.
Cost, at canon 8frame prices with a credit at $0.01 in a top-up pack: ElevenLabs TTS is 13 credits per 1,000 characters. A 150-word script is roughly 900 characters, so a full 60-second narration is one charge of 13 credits, about $0.13. That makes it cheap enough to generate five readings of the same script at different pacing and pick, which is a luxury the format never had.
What to cut
Every "in today's world" opener. The listener already knows what world they are in.
Adjective pairs. "Simple and intuitive" is one word doing the work of two.
The second example. If the first example landed, the second is padding. If it did not, a second will not save it.
Anything the picture already says.
An example, before and after
Before (68 words, 30 seconds, too long): "In today's fast-moving environment, teams are constantly looking for ways to simplify and streamline their creative workflows, which can often be complicated by the need to manage multiple tools and subscriptions across different platforms and vendors."
After (24 words): "Most teams run four tools to make one video. Four logins. Four bills. Four places for the file to go missing. There is a simpler way."
The second is shorter, concrete, and has a rhythm you can hear.
FAQ
How many words is a 30-second voiceover? About 75 at a comfortable pace, and 60 to 70 if the script needs room to breathe.
Should the voiceover be written before or after the visuals? Before, for explainers and anything narration-led. See how to storyboard an explainer video.
Is AI narration good enough for commercial work? For explainers, product videos, and social content, generally yes. For work where the voice is the brand, a human performance is still doing something synthetic reads do not.
Script, narration, and picture on one surface means you hear the timing against the cut instead of guessing. The 8frame canvas is free and unlimited, and generation is paid from $19/month.