Online ads are rarely won by production polish. They are won by how many versions you can test. Swapping one opening line or one voice can move CPA by 30% or more. With a human voice actor, every script change means a new booking and a new wait. With AI voiceover, you can go from revised copy to a shippable variant in minutes.
This guide breaks down the voiceover workflow for Meta and Google video ads using Soundwaver text to speech: ad specs, narration structure, a creative testing matrix, and the claims that get ads rejected.
Meta vs. Google: what changes for the voice
| Platform | Placement | Length | Narration strategy |
|---|---|---|---|
| Meta (Reels) | 9:16 vertical | 6–15s | Hook in the first 2 seconds, captions inside safe zones |
| Meta (Feed) | 1:1 / 4:5 | 15–30s | Pain point in 3 seconds, one proof point, one CTA |
| YouTube in-stream (skippable) | 16:9 | 15–30s | Earn the first 5 seconds, brand only after second 6 |
| YouTube Shorts | 9:16 | 15–30s | Faster pacing, narration cuts on caption beats |
| Google Demand Gen | Mixed ratios | 10–30s | Supply several headlines and voice reads for auto-assembly |
Rule of thumb: one master cut per placement. Do not write narration so generic that it fits every ratio — generate per format so the read matches the edit.
The 4-part structure of a converting ad read
[0–2s] Hook: name the viewer's exact situation
[3–8s] Amplify: say the pain they have not said out loud
[9–18s] Proof: one concrete number or one moment of use
[Last 3s] Single CTA: one action only
Example (15s, B2B tool):
“Still reconciling month-end across three spreadsheets? (hook) Miss one line and the whole report gets redone. (amplify) We help 400 small businesses close reconciliation in twelve minutes. (proof) Tap below to calculate your hours saved. (CTA)”
Writing notes:
- One idea per sentence. Listeners cannot re-read.
- Spell numbers for speech: write “twelve minutes,” not “12min,” so the model reads it correctly.
- Avoid absolutes: “guaranteed,” “#1,” and “cheapest” draw scrutiny on both Meta and Google.
Building a creative testing matrix with AI voice
Ad testing is about controlling variables. The real value of AI voiceover is that you can afford eight variants instead of two.
| Test dimension | Variants | Hold constant | Watch |
|---|---|---|---|
| Hook opening | 3 | Voice, visuals, length | 3-second view rate, CTR |
| Voice type (gender/age) | 3 | Script, visuals | CTR, CPA |
| Pace (fast/mid/slow) | 3 | Voice, script | Completion rate, CVR |
| CTA wording | 2 | Everything else | CVR, ROAS |
Cadence: one round per week, one dimension per round, and at least 3,000 impressions before you call a winner.
Five steps from script to live ad
- Pick the KPI first. Hook tests read on CTR; downstream tests read on CPA.
- Write three hooks. Same video, only the first two seconds and the opening caption change.
- Choose the voice. Finance and healthcare read steadier; consumer goods read warmer. Preview voices free on Soundwaver, and use WaveMind™ when you need more emotional range.
- Mix the read. Keep voice 6–10 dB above the music bed, and never let captions cover the product.
- Check compliance. Verify prices, discount windows, and any performance or income claim word by word before you spend.
Common mistakes
- Hooking with your brand name. “Hi, we’re Acme Tech” loses half the audience in two seconds.
- Three CTAs in one spot. Click, follow, comment — and nothing happens.
- Reading like a news anchor. Ads need the tone of one person talking to one person.
- Voice that fights the offer. A calm corporate read on a flash sale kills urgency.
- Never rotating creative. Four weeks on one asset means fatigue and rising cost.
Advanced: lock the brand sound with voice cloning
Across accounts, agencies, and markets, the risk is that every video sounds like a different company. Voice cloning keeps one approved brand voice consistent across all assets — with consent and per the voice cloning policy.
FAQ
Will platforms reject ads because the voice is AI?
Not for being AI-generated. Rejections usually come from misleading claims, fake before/after, or missing disclosures. Keep the read clear and do not imitate a real person.
How many versions should one ad have?
Six to eight: three hooks by two voices, or two hooks by three voices. Run small budgets to find the winner, then scale it.
How long should the script be?
Roughly 2.5–3 English words per second. About 40–45 words for 15 seconds, 80–90 for 30 seconds.
Do I need a commercial license?
Confirm your plan includes commercial usage — see Soundwaver pricing.
Next step
Take one live ad, rewrite only the first two seconds, and generate three variants with the same voice. Give each equal budget for 72 hours, then compare 3-second view rate and CTR.
