AI avatar video script formatter
Avatar voices read text very literally, so a script that looks fine on the page can sound rushed and robotic on screen. This formatter prepares your script for HeyGen or Synthesia: it inserts pause markers, flags sentences that are too long, surfaces proper nouns that may need a pronunciation guide, and estimates the finished duration.
Why avatar voices need different scripts than humans
When a human presenter reads a script, they instinctively add micro-pauses, vary pace, emphasise key words, and subtly slow down for transitions. AI avatar voices do none of this automatically — they read at a constant pace through punctuation marks and paragraph breaks, often sounding rushed and monotonous.
The standard fix is explicit formatting: pause markers that tell the synthesis engine to pause for a set duration, sentence structures that are short enough for the engine to handle as discrete units, and phonetic guidance for names and acronyms that the engine would otherwise guess at.
Raw prose written for reading becomes robotic narration. Prose reformatted for avatar delivery — shorter sentences, explicit pauses, phonetic annotations — sounds much closer to natural speech.
How it works
The tool processes your text in a few passes:
- Pause insertion — it adds a break marker (
<break>for HeyGen, or a pause hint for Synthesia) between paragraphs and after sentence-final punctuation so delivery has natural rhythm. - Sentence-length check — any sentence over ~25 words is flagged, because long runs make avatars sound breathless.
- Proper-noun scan — capitalised words that appear mid-sentence are listed as pronunciation candidates; these are where TTS most often slips.
- Duration estimate — word count ÷ ~140 wpm gives an approximate runtime.
Platform differences: HeyGen vs Synthesia
Both platforms use text-to-speech synthesis for their avatar narration, but they handle formatting differently:
HeyGen accepts [break] pause markers directly in the script input and
supports phonetic annotations for pronunciation. Paragraph breaks create natural
pauses without additional markers, but sentence-level pauses benefit from explicit
tags.
Synthesia handles pauses primarily through punctuation and paragraph structure. Explicit timing markers are less consistent across Synthesia’s engine depending on the voice chosen. The approach here is to use punctuation-based rhythm (shorter sentences, more full stops) rather than tag-based pauses.
The formatter applies the correct approach for whichever platform you select.
Common proper-noun problems
The pronunciation scanner flags capitalised mid-sentence words because these are where TTS most reliably mispronounces. Common problem categories:
- Company and product names — “Figma”, “Canva”, “Vercel” (engines often default to the phonetically obvious reading, which is sometimes wrong)
- Acronyms — “API” may be read as one word rather than three letters; “SQL” varies between “sequel” and “S-Q-L” by engine
- People’s names — any name unusual in spelling or origin is a candidate
- Technical terms — domain jargon the engine has not seen frequently enough to have a reliable pronunciation for
For each flagged word, decide whether to add a phonetic spelling in parentheses the first time it appears, or to listen to the engine’s default and accept it.
Tips for natural avatar narration
- Write for the ear, not the page — short declarative sentences read best.
- Spell out tricky names phonetically the first time, e.g. “Gera (GEH-ra)”.
- Keep one idea per sentence so pauses land where the meaning breaks.
- Read the formatted output aloud before rendering — if you stumble, the avatar will too.
- Use the duration estimate to plan. At roughly 140 words per minute, a 60-second segment is about 140 words. Trim before rendering; re-renders cost time and credits.