Three script edits do more for documentary narration than any voice swap: shorter sentences, numbers written the way you'd say them, and punctuation that sets the pauses. Most AI narration that sounds read aloud is a script written for the eye and handed to a voice.
The edits below are the ones I'd make before spending a single point. They work with any voice, but they matter most for the slow, serious genre where a flat read is obvious. If you're after the tool itself, the documentary voiceover maker is the place to paste a script and hear it.
How to make AI narration sound natural
A reader can go back. A listener can't. That one fact explains most of the difference between a page and a narration track: written sentences stack clauses because the eye can hold them, and a voice reading that stack has to rush the middle or breathe in the wrong place.
So write for a single pass. Keep one idea in each sentence, and put the important word near the end, where a voice naturally lands its weight. Let a long sentence run when the thought is building, then follow it with a short one. Ten sentences in a row at 18 words each will make any narrator drone. A workable mix: most sentences between 12 and 20 words, a beat of 4 to 8 words every few lines, and one in ten allowed to stretch to 30 when the thought needs the room. Count one page of your script against that mix and see where it drifts toward all-medium.
The expedition, which had been delayed by weather for nearly three weeks and which had lost two of its original members to illness before it ever left the coast, finally reached the pass in late October.
The expedition was late. Three weeks of bad weather. Two members gone before they left the coast, both sick. In late October, they reached the pass.
Don't take that too far. A whole documentary of fragments reads like a trailer. The goal is variety: one short beat for every two or three long sentences, and your hard facts at the ends of lines.
Numbers, dates and names written for the ear
Figures are where narration scripts betray their origin. Digits are compact on a page and awkward out loud, and a string like 1,847 or 3.2 million asks the voice to guess how you'd say it. Write what you want heard.
Years are the classic case. Spell the year the way a narrator would say it, round the number that nobody needs to the digit, and cut the ones the listener won't hold onto. Three figures in one sentence is a spreadsheet, and the ear gives up after two.
In 1847, the company reported revenues of $3,214,880, a 23.6% increase over 1846, across 41 stations.
In eighteen forty-seven the company took in a little over three million dollars. Nearly a quarter more than the year before. All of it from forty-one stations.
Names deserve the same suspicion. If a place or a person's name comes out wrong on the first listen, fix it before you render the whole script rather than after; the walkthrough on fixing a mispronounced name covers the options.
Punctuation and paragraph breaks set the pauses
You control more pacing through punctuation than through any setting. Commas give a short beat, periods give a full stop, and a blank line between paragraphs gives a longer breath. Use the blank line where a scene changes, and let ordinary sentences carry the rest.
When you need a pause of an exact length, the Studio model reads a break tag written inline. You can put up to 20 in one request, and they never show up in captions or SRT files. Save them for the moments that earn it: the beat before a reveal, or the silence after a hard fact.
The telegram arrived on a Tuesday. <break time="1.2s" /> It was three words long.
Nobody in the office spoke.Capitalizing a single word adds stress to it. WE had been warned reads differently from we had been warned. Use it once a page, not once a paragraph, or every line starts shouting.
Directions the Studio model follows
On the Studio model, which is the default on every plan, Free included, you can describe the delivery in plain English inside square brackets at the start of a line. Documentary work mostly wants restraint, so keep the instructions plain.
Write them in lowercase, in English even if the narration isn't, and give each line one coherent instruction. Two directions that pull against each other, a hush and a shout, make the result unpredictable. The voice emotion control page has more examples to borrow.
Pace
[slow and deliberate] at the head of a heavy paragraph. Use [very fast] almost never.
Tone
[calm and reassuring] or [articulate clearly at a steady pace] for explanatory stretches.
Volume
[very quiet] for an aside or a closing line. Match it to the words or it fights them.
Two limits worth knowing. A line that holds only a direction and no words is rejected, so every bracket needs something to speak after it. And on VocalLab Studio Flash, the faster, cheaper model, directions may be read out literally rather than performed, so check the model in Voice settings before blaming the script.
Set Delivery to Stable for long scripts
The Studio model has a Delivery setting with three options: Stable, Balanced and Creative. For documentary, I'd start on Stable. You want line 80 to sound like line 1, and a more varied delivery is exactly what you don't want over a forty-minute script. Move toward Creative only for short, dramatic pieces, and test one paragraph first.
Documentary voiceover script tips: a pre-render checklist
- Read it out loud once. Wherever you run out of breath, split the sentence.
- Mark every number. Spell years and rewrite any figure the listener can't hold onto.
- Check the names. Generate the hardest ones alone before the full pass.
- Set Delivery. Stable for the long haul, then change it only deliberately.
- Voice a scene at a time. Every regeneration is charged again at about 1 point per second, so fix a paragraph, not the whole film.
What a documentary narration costs in points
In the app, a script costs the character count divided by 17, rounded up, and one point buys about one second. A 10-minute narration is roughly 600 points, and a 20-minute one is roughly 1,200. Lite's 3,000 points a month covers about 50 minutes of finished audio, and Pro's 12,000 covers about 200.
Edits change that bill in your favor. Cutting a 1,700-character paragraph to 1,200 saves ⌈500 ÷ 17⌉ = 30 points every time that paragraph is generated. And because regenerating a 90-second scene costs about 90 points, it pays to settle the script on paper, then voice it once.
Voices with the weight for documentary work
Script edits can't turn a bright voice into a grave one. Two voices in the library start in the right register: a deep, time-worn American voice for serious or dark material, and an authoritative British voice that suits history and educational essays.
Why can't you just make a voice deeper? There's no pitch slider, and this post on deeper AI voices explains what to do instead.
Gravelly Male Documentary Voice
Serious documentaries and dark history
- American
- Deep, gravelly, time-worn
Authoritative British Male Documentary Voice
Historical documentaries and educational essays
- British
- Profound, authoritative
Frequently asked questions
Why does my AI documentary narration sound robotic?
Usually because the script was written to be read, not heard: long clause-stacked sentences, raw digits and no pause markers. Shorten sentences, spell out numbers and dates, and let punctuation and paragraph breaks set the rhythm.
How do I add a dramatic pause to AI narration?
Use a full stop or a blank line for natural pauses, or a break tag such as <break time="1.5s" /> for an exact length. The Studio model accepts up to 20 break tags per request, and they never appear in captions.
Can I tell the AI voice to sound slower or quieter?
Yes, on the Studio model. Put a bracketed instruction at the start of the line, for example [slow and deliberate] or [very quiet], followed by the words to speak. Studio Flash may read the bracket aloud instead.
Which Delivery setting is best for long narration?
Stable is the sensible starting point for long scripts because it keeps the read consistent from the first line to the last. Try Balanced or Creative on a single paragraph before using them on a whole project.