Natural voice-over: accents, numbers, short lines

Last updated 13 Oct 2026

Whether an AI voice (TTS or a talking character) sounds good depends mostly on how you write the lines.

Course · Stage 4: Scripts that hold attention · Lesson 14/40 · 8 min

Why it matters

AI voices often misread symbols ("710ml", "$25", "6–10am") and stumble on long, multi-clause sentences.

Regional words used in the wrong place sound fake, and local viewers notice instantly.

Step by step

1. Write numbers as words for the voice: "seven hundred ten milliliters", "twenty-five dollars", "six in the morning".
2. One idea per sentence, about 15 words max; an 8-second scene gets two short sentences.
3. Pick one accent for the whole video and use local words only where they fit naturally.
4. Read it aloud once with a timer: over 8 seconds → cut words, don't speed up the voice.
5. Listen to the AI read, mark mispronounced words and rewrite them the way they're spoken.

Common mistakes

• Leaving unit symbols for the AI to guess.
• One 30-word sentence carrying three ideas.
• Sprinkling slang just to sound local.

Ready-made prompt (ChatGPT / Claude / Gemini)

Voice-over for a [length] video: [paste lines]. Accent: [accent]. Rewrite it: numbers and units spelled the way they're spoken, max 15 words per sentence, max 2 sentences per 8-second scene, keep local words only where natural. Mark what you changed.

Exercise

Rewrite the voice-over of an old video with these 5 steps, then let an AI voice read it and listen.

Done when: No number or unit symbols left; no sentence over 15 words; the AI reads it without stumbling.

Need help? ngocdiep3135@gmail.com