Which voice sells? A small test for choosing an ad voiceover

A voiceover is one of the few parts of an ad you can change without reshooting anything. That makes it a good test variable. It also makes it easy to waste money, because it is tempting to pick the voice you like and call it a day. Your taste is not your customer's.

This page is for an owner or agency running paid ads who wants a measured voice decision. It covers the cheapest start, what little evidence exists, a blind listening test, a small paid test with pass or fail rules, credit maths, and the rights questions around AI voices. I also include a simple rule for when a phone recording beats a text-to-speech voice.

This page may contain affiliate links from which Blur Studio earns commissions.

The problem: you do not know which voice sells

A script can be strong and still fall flat in the wrong voice. Too salesy, too flat, too young, too slow. Hiring three voice actors to find out is a real cost, and the answer only holds for that script. So the goal is a cheap way to drop the obvious losers, then spend media money on a fair test between the survivors.

Cheapest first: record two readings yourself

Before any tool, record the script twice on a phone in a quiet room. One reading in your normal voice. One with a colleague or friend. Put a blanket over your head or desk if the room echoes. You now have two human versions and a feel for the pace the script needs.

These recordings are not wasted. They become your human control in the later test.

Phone recording or text-to-speech: a decision rule

• Use a phone recording when the voice should sound like a real person from your business (a founder, a local owner), when you have one or two ads, when the script changes rarely, or when you want to avoid any voice-clone consent question.
• Use text-to-speech (TTS), a tool that turns typed text into spoken audio, when you need several versions fast, when the script will change between tests, or when you need languages or accents you cannot record.
• Use both when you are testing. Keep the human reading as the control against which the AI voices must win.

What the evidence says (small and mixed)

I could not find an independent study of voice choice in ordinary commercial ads. Vendors that sell voice tools say voice testing improves performance, but they show no data, so I treat that as marketing.

The one controlled test I found is a 2023 political ad test run by Trilogy Interactive and described in an interview with its partner. It used a 15 second scripted ad read by four voices: a human woman, a human man, an AI woman and an AI man. According to that interview, the human female voice moved independent voters by 20 points, while the AI female voice performed very badly even though the tester thought it sounded the most human. The panel was political persuasion, not sales, and the voices were 2023 models, so I would not apply the numbers to your ads.

The lesson I take from it is narrow. Voice is a variable that can hurt as well as help, "sounds human" is not the same as "works," and a human control belongs in every test.

Your options for making the voices

Option	What it is	What to know
Yourself or a colleague	A phone recording	Free, human, limited in variety
A freelance voice actor	A person recorded to your script	Ask for a quote. Best when a human control beat AI, or the brand voice must stay consistent for years
ElevenLabs	A company that sells AI voices, with a large voice library	Starter $6 a month (30,000 credits, commercial license). Creator $22 a month (121,000 credits). The page also showed first-month offers with an unclear end date. The help center says 1 credit per character in the app. Confirm on the vendor page
Higgsfield Audio	One part of a website that bundles AI models	Text to Speech, Voice Change, custom voice from a sample. Cost shows on the button. See below
Voices built into the ad platform	Google Ads lets you add a text-to-speech voice-over to YouTube video ads from its asset library	Google's help post (July 2022) says it offers 7 English (US) voices, 4 male and 3 female, plus voices in nine other languages, and that you can try different voices with Video experiments. Its Sony result (25% higher ad recall) is Google's own claim from 2021 data. Check the current list in your account
Bulk ad voice tools	Tools built to render many hooks and calls to action at once	Worth a look if you need dozens of variants. Check claims; vendor pages show no lift data

For a wider look at voice tools, see best AI voice generator for business.

Write the script for the ear first

A weak read is often a weak script. Vendor guides from Noiz and Mika Reyes, and the research firm Zappi, agree on a few habits:

• Put the offer in the first sentence. "Fifty percent off this week" before any warm-up line.
• Short lines. Noiz suggests under 10 words a line and one benefit a sentence. A 15 to 30 second read is about 40 to 75 words (Mika Reyes).
• Match tone to the message. Zappi says to use a serious tone for a serious message and a light one for a light message, and to avoid acronyms and technical terms.
• Do not speed up to fit. Noiz warns that a 15 second slot read too fast sounds desperate. Cut the words instead.
• Emotion notes on key lines only. Both Noiz and Mika Reyes say over-tagging makes a voice sound acted. Tag the offer line, not every line.

Remember that these are vendor and tool-maker tips, not test results. Treat them as starting points and let your own test decide.

Step one: a blind listening test before you spend on media

1. Write one script of 20 to 40 seconds, written for the ear. Keep the call to action (CTA) as its own sentence at the end.
2. Pick three voice styles, for example warm and conversational, energetic, and calm and authoritative. Add your human recording as a fourth.
3. Make each version and name the files A, B, C and D. Do not tell listeners which is which.
4. Ask five people from your target audience to rank them and mark any that sound robotic, salesy or hard to trust.
5. Drop any voice two or more people flag. Keep two or three.

This costs almost nothing and removes the obvious losers. Check pronunciation as you go: brand names, prices and local place names. For Arabic or Hindi audiences, ask a native speaker to rate each voice. Accent fit often matters more than gender or "warmth."

Step two: one small paid test

Now spend a little on media, but follow the rules for a fair test. Meta's help page on A/B testing says to change one variable, choose a hypothesis first, use the same budget for each version, and avoid informal testing by switching ad sets on and off by hand, because audiences overlap and results become unreliable.

• Same creative, same script, same music, same length. Only the voice changes.
• Same budget per version, and use the platform's own test tool rather than turning ads on and off yourself.
• Pick the winning metric before you start, such as cost per lead. Watch hook rate (how many people keep watching past the first few seconds) and click-through rate (CTR, the share of viewers who click) too, but do not choose the winner by how much you like a voice.
• Set a minimum number of results per version in advance, for example 30 each.

I could not confirm a minimum run time from the Meta page I read, so I will not give you a day count. Run it long enough to reach the number of results you set in advance.

Illustrative pass rule (made up): call a winner only if one voice has at least 20 percent lower cost per result than the others and every version reached at least 30 results. Otherwise call it a tie and choose on brand fit. A tie is a useful answer: it means voice is not your problem, so test the opening three seconds instead. See ad hook tests and how to test 20 ad variations.

A trick that saves effort: keep the CTA as a separate audio clip. Then you can test "Book today" against "Learn more" without regenerating the whole read.

Mix and export so loudness does not decide the winner

Export every version at the same loudness. Noiz suggests about -14 LUFS (a standard loudness measure) for social feeds and keeping the voice about 6 dB above the music; those are one vendor's numbers, so check your platform's current audio specs. The point for a test is simple: if version B is louder than version A, you may be measuring volume, not voice. Name the files by voice, for example SummerSale-Warm-V2, so your media buyer can swap them without reopening the project.

Credit and cost maths with illustrative numbers

A 150 word script is roughly 900 characters. Three AI voices is about 2,700 characters. At 1 credit per character in the ElevenLabs app, that is about 2,700 credits, against 30,000 on Starter. So making the voices is cheap. The cost is media spend and your time.

For Higgsfield, the audio cost is not published; it shows on the Generate button. Make one 30 second read, note the credits, and multiply by the number of versions. Plans start at $19 a month for Starter (270 credits), with Plus at $59 ($47 on annual billing, 1,200 credits) and Ultra at $129 ($99 annual, 3,000 credits), as of November 2026. Check higgsfield.ai/pricing for current numbers. For more on how credits behave, see spending Higgsfield credits wisely.

Rights, consent and disclosure

• Use voices you have the right to use. Stock or library voices with commercial terms, or your own voice, or a cloned voice with written consent. Never use a sound-alike of a real person, and never a celebrity-style voice.
• Get a short signed release for any cloned voice used in paid ads, naming ad use and a duration.
• Higgsfield's help center says to clone only your own voice or one you have permission to use.
• Marketing is a commercial use. Check the plan's terms. A Free Higgsfield account adds a watermark, the pricing table lists commercial use as "Not included" for Free, and the help center says outputs are not restricted to paid plans. For ads, use a paid plan. For ElevenLabs, see its commercial use rules and my guide on voice cloning.
• Training terms. Higgsfield's help center, summarising its Terms of Use, says it may use content to train and improve its models except under Enterprise agreements. If a script contains confidential product details, think before you upload it.
• Disclosure. Some platforms label AI-generated content, and some laws require disclosure. This is information, not legal advice, so check local law and each platform's current rules. Do not present a synthetic voice as a real customer giving a testimonial. The US Federal Trade Commission (FTC) announced a final rule in August 2024 banning fake reviews and testimonials, and said it aims to deter AI-generated fake reviews. A made-up customer voice is the kind of thing to avoid, so check the current rule.

For a tool-by-tool view of cloning and consent, see best AI voice cloning tools.

Where Higgsfield fits and where it does not

According to Higgsfield's help center (August 2026), Audio offers Text to Speech with preset voices, emotion sliders and a choice of audio models, a custom voice made from a sample, and Voice Change on an existing recording. That is handy when your ad video is already being built on Higgsfield and you want a few voice versions in one place.

It is not an ad testing tool. It has no variant tracking and no way to measure which voice performs. The credit cost, the plan gating and the language list for audio are not published. If your main need is a deep voice library and a clear commercial license, ElevenLabs is the more focused choice. If you need one brand voice for years, a human actor is better.

Review checklist for each voice

• Brand name, price and place names pronounced correctly.
• Pace leaves room for the key claim to land.
• Emotion fits the claim (calm for trust, upbeat for offers).
• No robotic rhythm on numbers or lists.
• Audio level matches across versions, so loudness does not sway the result.
• Rights note saved: where the voice came from and what the license says.

Who should skip voice testing

• You spend very little on ads. You may not reach enough results to see a difference. Pick on brand fit and move on.
• Your hook or offer is weak. Fix that first; a better voice will not save it.
• You are a local business whose owner is the brand. Record yourself.
• You cannot get the rights or consent for the voices you want to test.

How I'd approach this

I would record two human versions, make two or three AI versions, and run the blind listening test with five people. I would keep the human reading as the control, and make the AI versions on whichever tool you already pay for, such as a Higgsfield paid plan or ElevenLabs. Then I would run one paid test with one variable, equal budgets and a pass rule written down before launch. If the result is a tie, I would take that as a sign to test something else.

If you run this often, the useful thing to build is the loop: scripts in one place, versions named clearly, rights notes saved and results logged. Blur Studio builds that kind of repeatable creative workflow, approval portal or shared asset library, which you can read about here. If you also work in several languages, see dubbing videos into other languages.