The voice is the easy part. What separates a convincing AI video voiceover from an obviously synthetic one is the script: short sentences, contractions, one idea at a time, and roughly 140 spoken words per minute of finished video. Audition voices on a sentence whose meaning depends on emphasis, script to the picture rather than the page, lead each visual by about a quarter of a second, and normalise the finished mix to around -14 LUFS for YouTube and social.
An AI video voiceover is narration written for your footage and spoken by a text-to-speech model rather than a person in a booth. The voices got good enough somewhere around 2024, which means the interesting problems moved: what makes a narrated video work now is the script, the pacing, and the sync — not the synthesis.
Most guides stop at “pick a voice, paste your script, download the MP3.” That is the part nobody struggles with.
This guide covers the rest: auditioning a voice properly, writing words that survive being spoken, scripting against picture, fixing a name the model keeps mangling, and what to normalise the finished mix to before you upload.
If you are still choosing software, two sibling posts do that job — the four best AI video voiceover generators ranked for paid tools, and six free AI video voiceover tools compared for the free tiers and their real ceilings. This one is about the craft, whichever tool you land on.
What an AI video voiceover actually replaces
An AI voiceover replaces three separate jobs: the performer with a microphone and a treated room, the writer who turns a feature list into spoken sentences, and the re-record cycle you lose an hour to when one line changes.
Most tools only replace the first. You still write, you still time it, and you still stitch the audio back to the picture yourself.
That is the biggest fork in this category. A text-to-speech box wants a finished script; a tool that watches the footage first will draft one for you. Both are covered further down.
How to judge an AI voice: prosody, pacing, breath, emphasis
Voice quality is not timbre. Almost every commercial model now sounds like a plausible human on a single sentence, so a five-word sample tells you nothing.
What separates a good model from a bad one is prosody — the pitch, duration, and energy the model spreads across a whole sentence. Four things to listen for:
Emphasis on the right word. Feed the model a sentence where the meaning depends on stress. “I never said she took the money” means six different things depending on which word is hit. A good model picks up the contrast from context; a weak one hits the verb every time.
Pitch reset between sentences. A human starts each new sentence a little higher and falls through it. Models that ignore paragraph structure produce a flat, unbroken shelf of speech that gets exhausting after ninety seconds.
Breath and pause. Real narrators breathe, and the breath is information — it tells the listener a clause just ended. Models that never breathe sound uncanny in long-form; models that insert breaths in the wrong place sound worse.
Pacing under a list. Lists are where synthetic speech falls apart. “Upload, translate, and export” should have a rising, rising, falling shape. If all three land identically, the voice will not survive a tutorial.
Build one audition script — 60 to 90 seconds — and run it through every voice you are considering. Include a question, a list of three, a number with a currency, your product name, a contrastive pair, and one deliberately long sentence.
Then listen to all of them with your eyes shut.
How to choose a voice for the content and the audience
Pick the voice for what the video is doing, not for which sample sounded nicest in isolation.
| Content type | What the voice needs | What to avoid |
|---|---|---|
| Software demo / tutorial | Mid-range pitch, unhurried, very clear consonants | Character voices, heavy vocal fry, anything “cinematic” |
| Short-form social | High energy, faster read, wide pitch movement | Flat corporate reads; they die in the first two seconds |
| Explainer / documentary | Lower pitch, slower, comfortable with long pauses | Upward inflection at the end of statements |
| Course / training | Warm, steady, consistent across hours of material | Anything distinctive enough to become annoying by lesson six |
| Ads and promos | Personality, and a read that can land a call to action | Voices that sound great narrating but weak selling |
Three rules that matter more than the table:
Match the accent to the audience, not to yourself. If most of your viewers are in the US, a US-accented read reduces friction even if you are not American. The exception is when your own accent is part of the brand — in which case the case for cloning your own voice is strong.
Consistency beats the perfect take. Pick one voice per channel or product and stay on it. Viewers build a relationship with a voice faster than with a face, and swapping it between videos costs you more than any single voice’s quality gains.
Audition at your real length. A voice that is charming for fifteen seconds can be unbearable at eight minutes. If you make long videos, test on a long passage.
Write narration for the ear, not the page
This is where most AI voiceovers are lost, and it has nothing to do with the model. Written English and spoken English are different languages, and a text-to-speech engine will read your prose exactly as written — including the parts that only work visually.
Keep sentences to 12 to 18 words. Anything over 25 loses the listener, and it also gives the model more chances to misplace the stress.
Use contractions. “You’ll see the export dialog” is speech. “You will see the export dialogue” is a manual.
Put the subject and verb at the front. “Although the file is large, the upload finishes quickly” makes the model carry an unresolved clause all the way through. “The file is large. The upload still finishes quickly.” does not.
One idea per sentence. If you need a semicolon, you need a full stop.
Delete everything that only exists on the page. Parentheses, “the former”, “see below”, “as mentioned above”, bulleted fragments, and em-dash asides all read badly out loud. A listener cannot scroll back, so useful repetition is a feature, not redundancy.
Say numbers the way you say them. Write “twenty-five dollars a month”, not “$25/mo”. Write “version two”, not “v2”.
Signpost. “Three things happen here. First…” gives a listener a structure they cannot see.
Then do the one test that catches everything else: read the script aloud at your normal speaking pace. Every place you run out of breath is a sentence to cut. Every place you stumble is a place the model will stumble.
Script to picture: writing against footage you already shot
Narration has a fixed exchange rate with time, and once you know it, scripting stops being guesswork.
Tools for Clear Speech at CUNY Baruch College, citing the National Center for Voice and Speech, puts the average English speaker at about 150 words per minute and recommends around 140 for comprehension. Narration over moving picture wants to sit at the slower end, because the viewer is reading the screen at the same time.
Budget 140 words per finished minute, then leave 10 to 15 percent of the runtime silent:
| Video length | Words to write | Roughly |
|---|---|---|
| 30 seconds | 60–70 | One short paragraph |
| 60 seconds | 125–140 | Two paragraphs |
| 90 seconds | 190–210 | Three paragraphs |
| 3 minutes | 380–420 | A page |
| 10 minutes | 1,250–1,400 | Three pages |
Three habits make scripting to picture much faster:
Write per shot, not per paragraph. Log the timecodes of the moments that matter first — the click, the chart appearing, the result loading — then write a line for each. You will find the video already has a structure.
Do not narrate what is legible on screen. If the button says “Export”, saying “click the Export button” wastes a second the viewer already spent reading it. Say why instead.
Leave air. A voiceover that runs continuously for three minutes reads as a machine even when the voice is flawless. Silence over a visual beat is the cheapest humanising trick there is.
Sync the audio to the visuals, and pace it to the cuts
Once you have audio, the job is alignment, and the direction matters.
If the footage is locked, you are fitting words to picture — cut the script, not the frames. If the footage is not shot yet, generate the voiceover first and edit the picture to the audio, which is far easier and is why explainer studios work that way.
Four alignment rules worth internalising:
Lead the visual by about a quarter of a second. Say the noun a beat before the thing appears, not after. Audio landing late reads as a mistake; audio landing marginally early reads as intentional.
Cut on the breath, not mid-word. Every pause in the generated audio is a legal edit point. Find them on the waveform and hang your cuts there.
Never straddle a hard cut with one clause. If a sentence has to span a cut, put the cut at a comma.
Fix overruns in this order. Trim words first. Then hold a frame or slow the picture. Only then speed the voice up — and stop at about 1.1x, because time-stretched synthetic speech develops a distinctive metallic edge well before recorded speech does.
For the mix: duck music 12 to 18 dB under the narration, or sidechain it to the voice track. Voiceover competing with a full-volume music bed is the most common reason a technically fine video sounds amateur.
And add captions. Verizon Media and Publicis Media’s 2019 study, reported by Next TV, found 83 percent of consumers watch mobile video with the sound off, and 80 percent said captions made them more likely to watch a video to the end.
Your voiceover is doing nothing for most mobile viewers until it is also on screen.
Pronunciation control: names, acronyms, and numbers
Every model will mangle your product name at least once, and this is the most fixable problem in the entire workflow.
The proper mechanism is SSML, the W3C Speech Synthesis Markup Language. The four elements that earn their keep:
<phoneme ph="…" alphabet="ipa">— give the exact pronunciation as an IPA string. The precise fix.<sub alias="…">— speak one thing, display another. Good for logos, symbols, and stylised spellings.<say-as interpret-as="…">— force a reading for dates, currency, telephone numbers, and spelled-out acronyms.<break time="400ms">— insert a pause the punctuation did not earn.
Not every consumer tool exposes SSML. When it does not, the workaround is reliable: respell the word phonetically in the script itself, generate the audio, then fix the spelling in your captions and on-screen text afterwards.
“Vidwip” becomes “vid-wip”. “Chargebee” becomes “charj-bee”.
Three more that catch people out:
Acronyms need a decision, once. Is it “S-Q-L” or “sequel”? “S-D-K” or “sudduck”? Write the choice into your style guide and force it every time, because models flip between the two.
Numbers and versions. Spell out “version two”, “two thousand and twenty-six”, “twenty-five dollars a month”. Read a URL as “vidwip dot com”.
Test names in a sentence. A word alone gets a different stress pattern than the same word mid-clause. If it sounds right in isolation and wrong in the video, that is why.
Keep a short pronunciation list per brand — ten to twenty entries — and paste it into every script. It takes ten minutes once and removes the single most common re-generation.
Loudness: the targets that decide how your video sounds
Loudness is measured in LUFS (Loudness Units relative to Full Scale) across the whole programme, not in peaks. Peak normalisation is what caused the problem loudness normalisation exists to fix.
Every distribution platform now normalises for you, which means being louder than the target buys you nothing and costs you dynamics. Match the target and keep your headroom:
| Destination | Integrated target | True peak ceiling | Source |
|---|---|---|---|
| YouTube | ≈ -14 LUFS (not officially published) | -1 dBTP | Measured via Stats for nerds |
| Spotify | -14 LUFS, per ITU-R BS.1770 | -1 dBTP (-2 if louder than -14) | Spotify for Artists |
| Apple Podcasts | ≈ -16 LKFS, ±1 dB | -1 dBFS | Apple Podcasts audio requirements |
| European broadcast | -23.0 LUFS, ±1.0 LU for live | -1 dBTP | EBU R 128 (2023) |
The published figures come straight from the specifications: Spotify’s loudness normalization page states -14 dB LUFS to the ITU 1770 standard and asks for true peak below -1 dBTP; Apple’s podcast audio requirements ask for roughly -16 dB LKFS with a ±1 dB tolerance; and EBU R 128 recommends normalising Programme Loudness to -23.0 LUFS with a maximum true peak of -1 dBTP.
YouTube is the one that is not published, and it is repeatedly measured at roughly -14 LUFS. You can check exactly what it did to your file: right-click the player, open Stats for nerds, and read the “Volume / Normalized” line — Production Advice explains the readout.
A value like “100% / 65%” means YouTube turned your audio down. It turns loud uploads down; it does not turn quiet ones up.
Two practical notes for AI narration specifically. Synthetic voice output is already unusually consistent — the loudness problem in an AI-narrated video is almost always the music bed or the screen-capture audio, not the voice. And a single ffmpeg pass gets you there without buying anything:
ffmpeg -i input.mp4 -af loudnorm=I=-14:TP=-1:LRA=11 -c:v copy output.mp4Run it two-pass if you care — loudnorm prints measured values on the first pass that you feed back into the second.
Where AI video voiceover still audibly fails
Being honest about the ceiling is what keeps you from shipping something that quietly undercuts you.
Sarcasm, irony, and jokes. Comic timing is a performance choice. No model reliably knows which beat to hold.
Sustained emotional range. A model can do warm, or serious, or excited. What it cannot yet do is move between them over twenty minutes the way a narrator naturally does, so long-form flattens.
Contrastive stress it cannot infer. “I didn’t say that” needs context the model does not have. If the emphasis carries the meaning, you will be fighting the output.
Code-mixed sentences. An English sentence containing a French product name, or a Hindi sentence with English technical terms, is where even strong multilingual models stumble on the switch.
Overlapping speakers. Crosstalk, interruption, two people finishing each other’s sentences — all still a scripted, sequential approximation.
Genuine hesitation. The half-second where someone searches for a word is a real signal of thinking, and inserting it artificially sounds exactly as artificial as it is.
Consistency across sessions. Regenerate one line six weeks later and the timbre can drift slightly from the surrounding audio. Regenerate a whole passage instead of a single sentence.
Anything where the person is the point. A founder’s story, an apology, a testimonial. If the reason to watch is that a specific human is speaking, a synthetic voice removes the reason.
Localisation is the opposite case — it is where synthesis clearly beats the alternative, because the alternative is booking a native speaker per market. The AI video dubbing tools comparison covers that side, including which tools preserve the original speaker’s voice across languages.
A faster path: let the AI watch the footage first
Everything above assumes you write the script. There is a shorter route for footage that already shows what it is about — screen recordings, walkthroughs, tutorials, product demos.
Vidwip’s AI Video Voiceover reads what is happening on screen, writes the narration scene by scene against it, and voices it. You start from narration already timed to your picture, rather than writing 400 words and then discovering they do not fit.
The specifics, so you can price it against the alternatives:
- 190+ studio-quality voices in 74 languages — the same library on the free plan as on paid.
- 60 credits per minute of finished video, and 1 credit is $0.01. That is $0.60 per finished minute. A three-minute walkthrough is $1.80.
- Free to start, no credit card — enough for a short narrated video.
- A small Vidwip badge on free-plan renders, and no watermark on paid plans. Free accounts get the same voices, languages, and models as paid ones.
- Steer the script before the render with a written brief, a voice and a language. It can’t be edited line by line afterwards, so a fix means a new render.
Paid plans add volume and output options rather than better voices: Starter is $25 a month for 2,500 credits and Pro is $99 for 10,000, and the paid tiers are where badge-free exports and batches of up to 5 videos (10 on Pro) live. They also run more renders at once: 4 on Starter, 10 on Pro, 1 on Free.
The fit is narrow, and worth stating plainly. If your video is a screen recording or a walkthrough, having the model watch first removes the hardest step; if it is b-roll cut to a script you have already written and rewritten, a plain text-to-speech tool from the generator comparison will serve you just as well.
The same engine produces the narrated walkthroughs covered in the AI product demo video maker roundup.
Where to go from here
The ranking of what actually determines whether an AI video voiceover works has been stable for a while now, and the voice is not at the top of it.
Script first — short sentences, contractions, one idea each, read aloud before you generate. Then pacing, at roughly 140 words per finished minute with silence deliberately left in.
Then sync, leading each visual by a quarter of a second and cutting on breaths. Then loudness, at around -14 LUFS with a -1 dBTP ceiling for YouTube and social.
Voice selection comes fifth. Pronunciation fixes are a ten-minute list you write once.
Get those five right and almost any current model will pass. Get them wrong and the best voice on the market still sounds synthetic — because what gives it away was never the voice.