How to get an AI voiceover that sounds like a person
The voice is rarely the problem. Modern text-to-speech can read a line so well that nobody notices it was never spoken. It cannot, however, read a line that was never meant to be spoken, and nearly every script handed to it is one of those.
Play most AI-narrated videos with your eyes closed and you can tell within a sentence. Not because the voice is synthetic; the better voices are past that. Because the sentences are wrong. They are too long, they have no breath in them, they put the important word in the wrong place, and they contain things nobody would say aloud. The voice is reading a document, and it sounds like a document being read.
The fix is almost entirely on the page. A voice model reads what it is given, with the rhythm the punctuation implies, and a script written for the ear comes back sounding like someone talking.
Write it the way you would say it
Read your script aloud, once, before generating anything. Everywhere you stumble, the voice will stumble. Everywhere you run out of breath, the voice will run on without one, which is worse.
Sentences for the ear are short. Fifteen words is long. A written sentence with two subordinate clauses is a spoken sentence with a wrong emphasis in it, because the voice has to guess which clause matters and will guess the first. Break it: one idea, full stop, next idea.
Put the important word at the end. Speech stresses the end of a sentence naturally, so "it heats up in forty seconds" lands the number and "forty seconds is how long it takes to heat up" lands nothing. Writers move the payload to the front for skimmers. Listeners cannot skim; they hear the end.
Use contractions. "It is" and "do not" are what documents say; "it's" and "don't" are what people say, and a voice reading the uncontracted form sounds like a legal notice. The only reason to write "do not" is when you want the stress on not.
Punctuate for breath, not grammar
Punctuation is the only pacing control most voice models have, so use it as one.
A full stop is a breath. A comma is a short pause. A paragraph break is a longer settle, long enough to change subject. An ellipsis, where the model supports it, is a hesitation. Commas that grammar would not require are fine if the breath is right: "And then, nothing" reads differently from "And then nothing", and the first is usually what you meant.
What a voice cannot do is read your mind about emphasis. If a word must be stressed, give it a sentence of its own, or put it at the end, or set it off with a comma before and after. Capitals and bold do nothing; the voice reads the letters, not the formatting.
Spell it the way it is said
The voice reads your script as written, and a script is full of things that are written one way and said another.
Numbers: "forty seconds", not "40s". "Two thousand and six", or "twenty oh six", whichever you would say; the digits "2006" will be read one way or the other and you will not know which until you hear it. Prices with the currency said aloud. Phone numbers as groups of words.
Abbreviations: spell them out, or write them the way they are spoken. "A.I." with full stops gets the letters; "ai" may get a syllable. Brand names and proper nouns that the voice mispronounces can usually be respelled phonetically in the script, which is invisible to the viewer and fixes the problem every time; the captions can be corrected separately afterwards.
Units, symbols, URLs: all of them. "Percent" rather than the sign, "slash" or "dot" where you want them said and nothing where you do not.
Generate in pieces
A script in one run is a script you have to regenerate whole when one word is wrong. Split it into sections of a few sentences each and generate each as its own clip. A fix then re-reads one piece, and every piece can be moved on the timeline independently, which is what you want when the picture under one sentence turns out to need another second.
It also lets the voice settle. Most models are at their most natural a sentence or two into a passage and slightly over-careful on the very first words; a section that starts with a short, easy sentence sounds better than one that opens on the hardest line.
Listen to each piece at full volume before building on it. A mispronunciation, a wrong stress, a number read as digits: all cheaper to fix now than after the captions have been generated from it.
Your own voice, if the video needs one
For a voice that is yours rather than a stock one, clone it: thirty seconds read in a quiet room, once, and every line after that is said in your voice. Everything above still applies, perhaps more, because the uncanny moment with a cloned voice is hearing yourself read a sentence you would never say. Write it as you talk and the clone sounds like you on a good day. The narration guide covers the set-up.
Mix it like a recording
A generated voice arrives clean, which makes it sound unlike anything recorded in a room, and the last ten percent of naturalness is in the mix rather than the script.
Put it on its own lane and duck the music under it: the music dips while the voice speaks and recovers in the gaps, which is how every produced video sounds. Leave the gaps. A voice that starts the instant the previous sentence ends is the other half of the hostage-tape effect; a breath's pause between sections is where the listener catches up.
If the result is still slightly too perfect, a touch of room reverb from the Audio tab's presets puts the voice in a space, and that space is what the ear uses to decide a person was there. Keep it small. A voice in a cathedral is a different problem.
Captions after, not before
Captions are transcribed from the audio, so generate them after the voice is final and placed. A phonetic respelling in the script will be transcribed as the real word, because the voice said it correctly; fix any that slipped. Running captions on a draft voice means running them twice.
The order that works
- Read it aloud first. Short sentences, contractions, the important word at the end.
- Punctuate for breath: full stops to breathe, commas to pause, paragraphs to change subject.
- Spell everything as it is said: numbers, abbreviations, names, units.
- Generate in sections, listen to each, fix the one that is wrong.
- Clone your own voice if the video needs you, and write for it the same way.
- Duck the music, keep the gaps, add a little room.
- Captions last, from the finished voice.