AI Voiceover for Ads: Fix Pronunciation, Pacing, and Captions

If your AI voiceover sounds wrong, we'd check the words before changing the voice. A script that looks tidy on a product page can become awkward when it's read aloud: a model number sounds like a quantity, a brand name comes out differently each time, or the last sentence rushes past the one detail someone needed to hear.

Start with a separate spoken version of your approved copy. Write names, numbers, and units the way you want them said, listen to the difficult sentence, then check it inside the full ad. Keep the correct written spelling for captions. That gives us a focused repair instead of another complete video to review.

Let's work through a fictional desk-fan ad and build a brief you can reuse. These are our editing recommendations and illustrative scripts, not generated audio we've tested or campaign results.

First, work out which part sounds wrong

Try listening once without looking at the captions. If you need the text to understand a product name or the offer, that's the first thing we'd fix. Then watch again with the picture: a perfectly clear sentence can still explain the wrong moment in the demonstration.

We wouldn't change all of those at once. If the only failure is the brand name, keep the rest of the read as your reference while you try to fix that sentence. A new voice or a full rewrite can make it harder to tell whether the original problem was solved.

Keep approved copy and spoken copy beside each other

Imagine the product is the Aven F2 desk fan. For this example, the brand owner wants Aven pronounced “AY-ven.” Our fictional product facts are: three speed settings, USB-C power, a cable included, and a power adapter sold separately. We have real footage of the fan and its controls. We don't have evidence for a noise rating, cooling performance, or battery life.

Here's how we'd prepare the tricky parts before asking for narration:

“AY-ven eff two” is an audition aid, not a spelling to put on the product or a promise that every speech tool will interpret it correctly. If the brand already has a pronunciation recording, use that as the reference for the person reviewing the read. Check that the tool supports using an audio reference before uploading it.

Numbers need the same care. “F2” names this model; it doesn't mean two fans. And writing out a unit should preserve the unit, rather than turn a specification into a stronger claim.

ElevenLabs' text-to-speech guidance, checked October 26, 2026, describes text normalization and pronunciation controls whose support varies by model and tool. So we'd start with plain spoken wording, then check the documentation for the exact generator if it still needs a dictionary, phonetic notation, or another control. Don't assume a tag copied from one model will work in another, or in the app using it.

Give the narration less work to do

Here's a first script we might receive for the fan:

Meet Aven F2, the USB-C desk fan with three speed settings and a compact design
for your workspace, including a cable, with a power adapter sold separately,
available now on our product page.

It contains several details, but they're strung into one long sentence. We'd split the explanation and decide what deserves the viewer's attention:

Meet AY-ven eff two.
Choose from three speed settings.
Plug it into U S B C power. The cable is included;
the power adapter is sold separately.
See the fan on our product page.

The punctuation is a starting direction for the read. It isn't a timing specification. We'd listen for how this voice handles the breaks, especially between the included cable and the separately sold adapter. Those details belong together so the buyer understands the purchase.

If the ad must be shorter, we'd consider dropping the model name from the spoken opening while keeping it visible as product text. We could say “Choose your desk-fan speed,” show the real control, and continue with the power explanation. That's a proposed copy change to approve; it shouldn't happen silently during audio editing.

We wouldn't squeeze the line into a fixed duration by deleting “sold separately.” And we wouldn't pick a universal word count and assume it guarantees the runtime. Read the actual take against the footage. If the demonstration needs more time, shorten a secondary message or change the edit.

Audition the difficult sentence before assembling the whole ad

For this example, we'd start with “Meet AY-ven eff two” and the power sentence. Together they include the brand, model, abbreviation, and qualification that need the closest attention.

Keep the voice and language consistent while comparing a small wording change. Note the actual input and the settings the tool exposes, so you can return to the version you accepted. If you're using a person's voice or likeness, confirm permission for the intended use before creating it. A generic synthetic narrator can explain the product without impersonating a customer.

Once the pronunciation works in the short passage, listen to the complete narration. A word can sound fine in isolation and awkward beside the surrounding words. Check the joins too: replacing one phrase may leave a sudden change in tone, loudness, or background sound.

If your tool only regenerates longer passages, review that whole replacement. Local repair is our goal; it's something to verify in the output, rather than assume the tool preserved everything. If you can't get a clear, consistent read, simplify the wording or use a human narrator for the passage. More attempts aren't automatically a better use of the budget.

Captions should keep the real spelling

The spoken input might say “AY-ven eff two.” The caption should say Aven F2. We'd keep that mapping next to the script so a pronunciation workaround doesn't become a new spelling in the finished ad.

Captions still need to reflect what's actually heard. If a generated take says the adapter is included, a correct caption cannot repair the incorrect audio. Replace the take, then check both. The W3C caption guidance, checked October 26, 2026, explains that captions convey speech and relevant non-speech audio, and that automatic captions need accuracy review. That makes them more than decorative keywords over the video.

For the fan, we'd check the brand, model, USB-C, speed settings, and adapter sentence against the approved copy. Then we'd check when each caption appears. After an audio replacement, an old caption can arrive before the new line or disappear before the qualification finishes.

Keep extra product labels separate from captions. A label saying “3 speed settings” can help identify the control in the footage, but it doesn't replace captions for the rest of the narration. Watch at the size someone will see on a phone, with the platform preview when available, and make sure neither type of text covers the control you're demonstrating.

Ask for a repair that names what should survive

We build Cospark, so the product recommendation here is our own. The workflow also works with another editor and speech tool. Script to Video is our entry point for turning a script into narrated video, and Cospark's product page describes combining uploaded footage with voiceover, captions, and conversational edits. We checked those public descriptions on October 26, 2026; this guide doesn't test paid generation or promise a particular pronunciation control inside Cospark.

Bring the approved script, real fan footage, and the spoken/display mapping. Then make the correction specific. Here's a prompt we'd adapt after hearing a wrong brand name:

Repair the opening narration for the Aven F2 desk-fan ad.
The approved pronunciation is AY-ven eff two. Audition that sentence first.
This spelling is for speech direction only. Display Aven F2 in captions.

Keep the approved product footage, scene order, three-speed explanation,
and closing destination. Preserve the statement that the USB-C cable is
included and the power adapter is sold separately.

If the replacement narration changes duration, adjust the affected captions
and show me where the edit changes. Do not remove a qualification to make it fit.
Do not regenerate accepted product footage to fix an audio problem.

Show the revised passage with the surrounding audio, then the complete ad.
I need to check the pronunciation, joins, captions, and final mix.

That's a request to review, not a guarantee that every part will stay untouched. Compare the result with the version you accepted. If you'd like a broader starting brief, our first-project guide walks through choosing assets and planning the whole ad.

Listen once more after export

We'd review the narration alone, then the full mix, then watch with the sound off. Each pass has a different job: hear the words, check they survive the music and edits, and see whether the captions carry the explanation. The silent pass doesn't replace listening.

For our fan ad, we'd save a small approval note:

Approved product: Aven F2 desk fan
Voice / language / exposed settings: record the actual selection
Spoken input version: attach the accepted text
Brand reference: AY-ven; model F2 read as eff two
Written forms: Aven F2, USB-C
Must retain: three speeds; cable included; power adapter sold separately
Review: opening pronunciation, full read, joins, mix, captions, phone preview
Final file: record the exact exported filename and version reviewed

Open that exported file rather than approving only the editor preview. Listen through the first and last words, pause on the product text, and confirm the destination still matches the offer. Keep the rejected takes and actual charges in your production notes if you're evaluating cost; this example doesn't establish how many attempts your ad will need.

For the next version, we'd keep the accepted narration and change one thing with a clear purpose. If you're still choosing the message itself, the AI ad creative guide can help with that earlier decision. Once the words are approved, getting them said clearly—and keeping their meaning intact through captions and export—is a useful job of its own.