How to Make an AI Voice-Over (and the Licence Detail Most Guides Miss)

Text-to-speech crossed the line from robotic to convincing, and it did so recently enough that
most people’s impression of it is out of date. If you need narration for a video, a course, a
podcast intro or an audio version of an article, this is what the process actually involves.

How do you make an AI voice-over?

Paste your script into a text-to-speech tool, choose a voice, and download the audio file. There
is no recording, no microphone and no editing step. A one-minute narration takes under two minutes
end to end, and most of that is choosing the voice.

Write the script for the ear

The most common disappointment is a good tool reading a script written to be read silently. The
two are not the same.

• Short sentences. A sentence you can say in one breath sounds natural. One
  with three clauses does not.
• Punctuation is timing. A full stop is a pause; a comma is a beat. This is your
  main control over pacing, and it works better than any setting in the interface.
• Write numbers as words where the reading is ambiguous. “1990” might be read as
  a year or as one thousand nine hundred and ninety.
• Read it aloud once yourself. Anywhere you stumble, the model will too.

What it costs, and the licence trap

Text-to-speech is usually billed per character rather than per minute. ElevenLabs’ published
plans are a useful reference point:

Plan	Monthly credits	Commercial licence	Price
Free	10,000	No	$0
Starter	30,000	Yes	$6 / month

One credit is roughly one character, so 10,000 credits is about 10,000 characters — several
minutes of narration, and plenty to decide whether the output is good enough for you.

The line that matters most is the middle column. The free plan carries no commercial
licence. If the audio ends up in a monetised video, a client’s project or anything
promoting a business, the free tier is not the right plan — and this is the detail most guides
leave out.

Choosing a voice

Spend your time here rather than on settings. Three things to listen for, using the same
sentence from your own script for each candidate:

• Pace. Slower than feels natural in a preview usually sounds right over
  video.
• Warmth against authority. Explaining something benefits from warmth;
  instructions benefit from directness. Mismatched tone is what makes narration feel synthetic.
• Consistency across a long passage. Some voices hold up over a paragraph better
  than a single line suggests. Always audition with a real paragraph, never with a demo sentence.

Cloning your own voice

Most tools can build a voice from a sample of your own speech, which is worth it if you produce
regularly and want a consistent identity across everything you publish. It carries obligations a
stock voice does not: consent is required for any voice that is not yours, and impersonation is
prohibited by every provider’s terms. Our
guide to voice cloning tools covers what each one
requires.

Making it say the word properly

Names, acronyms and technical terms are where synthesis most often gets it wrong, and the fix is
in the script rather than in any setting.

• Spell it how it sounds. Writing a difficult name phonetically in the script
  usually fixes it outright. The listener hears the audio, not your spelling.
• Space out acronyms you want read letter by letter — “A P I” rather than
  “API” — and leave them closed up when you want them read as a word.
• Give units their full words. “GB” may be read as “gee bee”; “gigabytes”
  never is.
• Use a comma to slow a list. Three items separated by commas are read as
  three items; without them they run together.

Generate one short test line with the difficult words in it before committing a long script.
Ten seconds of checking saves regenerating five minutes of narration.

Where AI narration is the wrong choice

Worth saying plainly, because the tools will not:

• Anything emotionally heavy. Grief, apology, genuine excitement. Synthesis
  delivers competent neutrality, and neutrality in those contexts reads as indifference.
• Your own personal brand, if your audience knows your voice. A substitute is
  noticed.
• Long-form fiction and audiobooks. Consistency holds for minutes, not hours;
  the sameness becomes audible over a chapter.

For narration, explainers, courses, product walkthroughs and accessibility audio — the large
majority of what people need — it is genuinely good enough that the question stops being
interesting.

Frequently asked questions

Is AI voice-over free?

Free tiers exist and are usable — ElevenLabs gives 10,000 credits a month. What free generally
does not include is a commercial licence, which starts at $6 a month on their Starter plan.

Can people tell it is AI?

Usually not, on a well-written script. What gives it away is almost always the writing — long
sentences, no pauses, tone mismatched to the content — rather than the synthesis itself.

How many minutes is 10,000 credits?

Roughly 10,000 characters of speech. The exact duration depends on the voice and its pace, so
treat it as several minutes of narration rather than a fixed number.

Can I use an AI voice for a YouTube video?

Yes, subject to the licence on your plan. A monetised video is a commercial use, so it needs a
plan that grants commercial rights.

Can I clone someone else’s voice?

Not without their consent. Every major provider requires it and prohibits impersonation, and
this is one of the few rules in the field that is consistently enforced.

Next: use it in a faceless reel · clone your own voice instead