You need forty seconds of someone saying your business's name clearly: over a product video, on the phone while a customer waits, or in a spot a local station will air. A voice actor has to be found, booked and briefed. Your own recording sounds like you, in the back office.

Call 555-123-4567 or book at example.com/book.

A voice can read it as one long figure

Call five five five, one two three, four five six seven.

Or book at example dot com slash book.

Every number the way a person says it

Fig. 1The same line written for the eye and for the ear: the phone number in spoken groups, the web address as it's said. A sample number and address, in the prompt's own format.

An AI voice can read it for you. The usual failure is a script written for reading, read aloud: a phone number that comes out as one long figure, a street name said the way it's spelled, a sentence too long to breathe through. So write for the ear first, then generate.

Write the script for the ear

the step that pays off

Before you open anything, write down three things. Where the audio will run and how long it can be. Every fact it has to say: the offer, the phone number, the web address, the hours. And every name a stranger could get wrong, with how you say it.

Paste those into the prompt below, in Claude. It writes lines short enough to say in one breath, and it writes every number the way a person says it. That second part is the fix ElevenLabs' own guide gives for phone numbers, prices, dates and web addresses, which its voices can misread as written. Its example turns "123-456-7890" into "one two three, four five six, seven eight nine zero".

Then check the numbers list against your facts, figure by figure. A wrong digit in a phone number in a radio spot is a listener who rings someone else.

Step 1: Write the script for the ear

Write a voiceover script that a text-to-speech voice will read aloud.
Nobody will see it on a screen, so write it for the ear.
Rules:
- Every fact must come from the facts below. If the script needs a fact
  the facts do not give, write [ASK: the question] instead of guessing.
  Never invent an offer, a price, a date or a claim about the business.
- One idea per sentence. Keep every sentence short enough to say in one
  breath.
- Write every number the way a person says it out loud. Phone numbers in
  spoken groups ("five five five, one two three, four five six seven"),
  prices in words ("forty-nine dollars"), dates in words ("Saturday the
  fourteenth"), web addresses as spoken ("example dot com slash book").
  No digits and no symbols such as %, $, & or / anywhere in the script.
- Expand every abbreviation: "St." becomes "Street", "Dr." becomes
  "Doctor".
- Say the business name and the one thing the listener should do early,
  and repeat that one thing once at the end.
- Fit {{LENGTH_IN_SECONDS}} seconds read at an unhurried pace.
- Plain words. No exclamation marks. No questions like "Are you tired
  of".
Return three parts:
1. The script, one sentence per line.
2. Numbers check: every number, price, date, phone number and web
   address in the script, as it will be spoken, next to how it appears
   in the facts.
3. Names: every brand, product, person, street and place name in the
   script, with its pronunciation note from the names list. If a name a
   stranger could say wrong has no note, write
   [ASK: how do you say this name?].
What it is for and where it will run:
"""
{{WHAT_IT_IS_FOR_AND_WHERE_IT_RUNS}}
"""
Facts it has to say:
"""
{{FACTS}}
"""
Names and how you say them:
"""
{{NAMES_AND_HOW_YOU_SAY_THEM}}
"""

Checkpoint

A script in short spoken lines, a numbers list you've checked, and a names list with how each one sounds.

Pick a voice and make a first take
In ElevenLabs, open Text to Speech and paste the script. ElevenLabs ranks the voice as the setting that matters most, then the model. Pick one from your Voices.

For more choice, open the Voice Library. Filter by language, accent and quality ("Studio Quality" is one option), and play any voice's sample. Save the ones you like and they appear in the picker.

Run your first two lines through three voices before you commit to one. ElevenLabs also suggests trying a few models with the voice you pick, because some pair better than others.

ElevenLabs calls stability around 50 and similarity around 75 the most common setting, so start there, where the model shows one. Not every model has every slider: ElevenLabs lists which ones Eleven v3 and v4 leave out. Lower stability gives a wider range and can turn erratic. Higher gives steadier delivery and can sound monotone. For an on-hold message, steady wins.
Pro tip
Generation isn't deterministic, so the same line comes out a little different each time. Make two or three takes of any line that matters and keep the best one.
Checkpoint
One voice and one model you've heard read your own lines.
Teach it your names
Take the names list from step one and open a pronunciation dictionary. Add one rule per name. Rules are case-sensitive, so type each name exactly as the script spells it.

Start with an alias rule: it swaps in a respelling before the voice reads the word, the way ElevenLabs' guide spells "SaaS" as "sass". If the alias still sounds wrong, switch to a phoneme rule with an IPA spelling. Not every model takes phoneme rules, and the guide has a table of which ones do.

ElevenLabs' guide says a saved dictionary applies across every script you write in the Text to Speech app, so later scripts start with your names already set. If the listen-through in the next step flags a name, add a rule here and leave the script's spelling alone.
Checkpoint
Every name on your list said the way you say it.

Listen through once and fix what you flag

Play the whole take from start to end on the speaker your customer will hear it on: a phone for on-hold, a laptop for a product video. Keep the script in front of you and mark every line that:

says a number, a name or a web address differently from your lists
runs too fast, or stresses the wrong word
sounds odd next to the line before it

Time it too. If it runs long, cut a line. Leave the speed slider alone: ElevenLabs notes that extreme speed values can affect quality.

Paste the script and your flagged lines into the prompt below, put the new script back into Text to Speech, and generate again. Repeat until a full listen-through flags nothing.

Step 2: Rewrite the lines that sounded wrong

Below is a voiceover script and a list of lines that sounded wrong when
a text-to-speech voice read them, each with a note on what was wrong.
Rewrite only the flagged lines. Keep every other line exactly as it is.
Rules:
- Do not change any fact. A number, price, date, name or web address
  stays the same; only the way it is written for speaking may change.
- A line that ran too fast or ran out of breath: split it into two
  sentences.
- A line that stressed the wrong word: reorder the sentence so the word
  that matters comes last.
- A name that was said wrong: do not change how it is spelled in the
  script. Leave the line as it is and add the name to a list at the end,
  with a respelling of how it sounds (hyphens between syllables, the
  stressed syllable in capitals), for my pronunciation dictionary.
- If you cannot tell what was wrong from the note, leave the line and
  write [ASK: what did it sound like?] under it.
Return the full script, one sentence per line, with [CHANGED] at the
start of every line you rewrote. Then the list of names for the
dictionary, one per line: the name exactly as the script spells it,
then the respelling.
Script:
"""
{{SCRIPT}}
"""
Flagged lines and what was wrong:
"""
{{FLAGGED_LINES_AND_NOTES}}
"""

Checkpoint

A take you've heard start to end with no line flagged.

Clear the rights before it runs
Two checks before the file leaves your laptop.

The licence. ElevenLabs says audio made on the free plan, or while signed out, can't be used commercially. A paid plan is one of several conditions. The content must not come from its beta services, you need the rights to what you publish, and the use has to follow its Terms of Service and Prohibited Use Policy. A product video, an on-hold message and an ad are all the business talking, so make the final take on a paid plan. Download it from your history as MP3 or WAV. ElevenLabs says WAV files need to come from your history.

The platform. YouTube asks creators to disclose realistic AI content, including content that makes a real person appear to say or do something they didn't. Voice Library voices are shared Professional Voice Clones, which means they're real people's voices. A stock voice reading your script isn't named on either of YouTube's lists, so when in doubt, tick the disclosure. A radio station or a podcast host can have rules of its own, so ask before you send the file.
Checkpoint
A file made on a paid plan, checked against ElevenLabs' terms and the rules of where it will run.

The takeaway

You end up with a file in a voice you picked, saying every name and number the way you would, and two things that carry over to the next one: a script prompt that writes for the ear, and a pronunciation dictionary with your names in it.

The limit: this reads a script well. It doesn't perform one. Comedy timing, a character, or a read that has to react like a real person is still a voice actor's job. And if the script has a wrong phone number in it, the voice will say the wrong number clearly.

Variations

Your own voice, without recording every script. ElevenLabs' Instant Voice Cloning works from about one to two minutes of clean audio with no reverb or background noise, and asks you to confirm you have the right and consent to clone the voice. Record once in a quiet room, in the delivery you want, because the clone copies how you read. The Professional version is stricter still: it lets you clone only your own voice, verified by reading lines aloud. On YouTube, "cloning one's own voice to create voice overs or dubs" is on the list of changes creators don't have to disclose.

Never clone anyone else's voice without their written permission. Not a staff member who has left, not a local radio host, not a famous voice. For a Professional clone, ElevenLabs says no outright: "Even with their consent, you cannot clone someone else's voice."

An on-hold message that changes each season. Keep the facts, the voice and the dictionary. Change only the offer line and generate again. If the voice came from the Voice Library, check its notice period first: it's how long you keep the voice if its owner stops sharing it, and a voice with none goes straight away.

Don't run it in a second language unheard. ElevenLabs says a voice keeps its original accent in another language on most models, and that Eleven v4 speaks the target language fluently instead. Either way, have someone who speaks the language listen through before it goes out.

← All use cases

Make a voiceover without a studio

How it runs, start to finish.

Write the script for the ear

Pick a voice and make a first take

Teach it your names

Listen through once and fix what you flag

Clear the rights before it runs

The takeaway

Variations

Want this running in your business?