How to prepare a script for text-to-speech (with example)

Prepare a text-to-speech script by writing the words your audience should hear, resolving ambiguous dates and abbreviations, and separating narration from production notes. Test the difficult passages before generating the complete recording. A polished written paragraph is not automatically a clear spoken paragraph: listeners cannot see your punctuation, scan a table or reread a complicated condition at a glance.

Open Voice & Audio, pick ElevenLabs and paste the trickiest passage of your script.

For a whole document (PDF, Word or PowerPoint), Deepnia’s document-to-audio tool lets you validate the transcript before generating the audio.

Keep an approved source beside the spoken version

Create two documents or clearly labeled sections. The source holds the approved facts. The spoken version adapts their presentation for listening. Do not overwrite the source while editing: it is your reference when checking whether a helpful rewrite has accidentally changed the meaning.

For example, “booking recommended” must not become “booking required.” A smoother sentence is not an improvement if it changes what the audience must do. Likewise, adding “limited spaces” or urgency introduces claims that need evidence.

Write down the intended listener, language and destination of the audio. An instruction heard during a demonstration may rely on a visible object; an audio-only announcement cannot. Decide which information must be understandable without looking at a screen before you shorten anything.

Worked example: turn a notice into narration

Original notice:

Intro session: 11/06/2026, 2:30–4 p.m. Room C. Demo + Q&A. Booking recommended by 11/03. Bring a laptop if available. Details at reception.

First, resolve the dates. Does “11/06” mean November sixth or the eleventh of June? Do not let an AI guess. For this example, the organizer confirms November sixth, with a booking recommendation of November third, both in 2026.

Spoken version:

Join our introductory session on November sixth, twenty twenty-six, from two thirty to four in the afternoon. We will meet in Room C. After a demonstration, there will be time for questions and answers. We recommend booking by November third. If you have a laptop available, please bring it. Ask at reception for more information.

The rewrite makes the date order explicit, expands the question-and-answer abbreviation and gives the fragments a conversational structure. It preserves the optional laptop and recommended booking. It does not invent the subject of the demonstration or a promise about what participants will achieve.

For an audience accustomed to day-first dates, “the sixth of November” may be a better editorial choice. Both versions identify the same day. Choose a convention suited to the audience and keep it consistent across the recording, captions and event page.

Review numbers by meaning, not appearance

Numbers can represent quantities, years, measurements, codes or individual digits. A room number and a telephone number do not necessarily need the same reading pattern. Decide what the listener needs to recognize before expanding the text.

Spell out numbers, dates, symbols and acronyms whenever the intended reading is unclear. MiniMax says MiniMax Speech 2.6 HD reads dates, amounts and phone numbers in their usual form.

Build a small ambiguity list before your first test:

Written item	Question to answer	Possible spoken form
11/06	Which date convention applies?	November sixth, once confirmed
1.5 hours	Is a decimal necessary?	one and a half hours
Q&A	Will beginners understand the initials?	questions and answers
Code B14	How is the identifier normally spoken?	code B fourteen, if approved

Do not silently add missing information. An online event may need a time zone, but a script editor cannot determine it from “2:30 p.m.” alone. Flag the missing zone for the organizer rather than using the voice's apparent location or accent as a clue.

Remove instructions that are not meant to be spoken

Production notes such as “show the product,” “pause here” or “sound reassuring” belong outside the narration. Pasted into the text field, they may be read aloud.

Keep a clean narration block and a separate direction block. For a pause or a softer tone, Open TTS accepts tags such as [pause] or [soft], placed just before the passage they affect. Otherwise, start with ordinary sentences and punctuation.

Paragraph breaks make your script easier to review. If timing matters, measure the generated recording. Repeated ellipses are not a reliable substitute for a measured edit.

Ready to try your prepared text? Paste it into Deepnia’s voice-over studio and compare two voices, for example ElevenLabs and Gemini Flash TTS. Use a short passage containing your difficult words and numbers so the first test answers a useful question.

Rewrite for one-pass understanding

Read each sentence aloud yourself. If you run out of context before reaching its main point, move the point earlier or split the sentence. Replace long parenthetical explanations with a separate sentence when the detail is necessary. Remove references such as “see above” if the recording will stand alone.

For a process, state the action and then the condition: “Save the file before closing the window.” Avoid a chain of pronouns when several objects are involved. “Move it beside that” might be clear beside a diagram but confusing in an audio-only instruction.

Example prompt:

Rewrite the following source as an English narration script for beginners. Preserve every fact, exception and requirement. Add no benefits, urgency or claims. Expand ambiguous abbreviations only when their meaning is established. Flag uncertain dates and identifiers separately. Return a clean spoken script followed by a change log. Do not include stage directions in the narration or claim an exact audio duration.

Compare the result against the source rather than accepting it because it sounds natural. Check negatives, deadlines and optional actions first. “You do not need an account” and “you need an account” differ by a small phrase but describe opposite requirements.

Test pronunciation without losing the correct spelling

Select a passage containing a proper name, a number and a representative sentence. Listen once without reading. Then listen again while following the script, checking for missing words or changes in meaning.

If a name is wrong, confirm the intended pronunciation with an appropriate source or the person concerned. You can experiment with a spoken-input spelling if necessary, but retain the correct written name in your source, captions and public materials. A pronunciation workaround is not a new official spelling.

Change one variable at a time. Record the problematic passage, the edit you made and what you heard afterward. Changing the voice and rewriting the entire script together makes it harder to identify what helped.

Once the test works, review the complete recording. Check the opening, closing action and any passage revised after approval particularly carefully.

Frequently asked questions

Should I spell out every number?

Not automatically. Prioritize ambiguous items and anything the voice reads incorrectly. Keep the original numeric values nearby so you can check that an expanded script has not changed a quantity or identifier.

How many words fit into thirty seconds?

About 70 to 80 words at a typical narration pace. Pace, pauses and wording change the final duration, so generate and measure the audio. For a small adjustment, the Open TTS Speed setting runs from 0.70× to 1.20×; beyond that, remove secondary information instead of making essential instructions difficult to follow.

Should captions copy the voice input exactly?

Captions should accurately represent the speech, but they can use readable written forms such as numerals for dates. Do not copy deliberate pronunciation misspellings into the captions. Review their wording and timing against the final audio, not an earlier take.

Prepare one paragraph before scaling up

Choose a short source paragraph, resolve its ambiguities and create a spoken version. For the next stage, read about creating a natural AI voiceover or choosing a voice for your audience. For long content, see choosing an AI voice for long-form narration and building an AI narration workflow. Clear delivery starts with a script whose meaning survives being heard once.

Open Deepnia’s voice-over studio, pick ElevenLabs and paste your paragraph.