Writing narration scripts for slides that sound human instead of read aloud

The fastest way to make a converted video sound bad is to feed the slide’s own text to the voice. It produces something technically correct and instantly recognisable as a machine reading bullet points.

The reason is that slide text is written for the eye. Bullets are fragments, they have no verbs, and they rely on the viewer supplying the connections. Spoken language has to supply them.

Explain, do not repeat

A bullet that reads “Q3 revenue: up 12% YoY” is perfect on a slide and terrible as speech. Read aloud it becomes “Q3 revenue, up twelve percent, year over year” — three fragments in a row with nothing holding them together.

Spoken, the same content is: “Revenue grew twelve percent compared with the same quarter last year. That is the second consecutive quarter of growth.”

Same information, one extra clause, and it sounds like a person who understands the material. The general rule: the narration should explain the bullet, not read it.

Mechanics that make it sound spoken

Complete sentences with verbs. Fragments read as a list. Sentences read as speech.

Shorter than you would write. Written English tolerates clauses; spoken English does not. If a sentence needs a breath halfway through, split it.

Active voice. “We changed the approval flow” beats “the approval flow was changed”.

Signposting. “Three things matter here” then naming them. “Here is why this matters” before the conclusion. Speakers do this constantly and writers strip it out — put it back.

Contractions. “We’ll” and “don’t” sound like speech. “We will” and “do not” sound like a terms-of-service document. Unless the register is formal, contract.

The symbols that break text-to-speech

Text-to-speech handles prose well and symbols unpredictably. Before you render, replace:

• “12%” → “twelve percent” (some engines say “twelve percent”, some say “one two percent”)
• “Q3” → “the third quarter” or “Q three”, depending on which reads better aloud
• “YoY” / “MoM” / other acronyms → spell out the first time
• “1,200” → “one thousand two hundred” if the engine reads it as “one point two”
• ”→” and ”&” → “leads to” and “and”
• URLs → either remove them or write “see the link in the description”; nobody wants a URL read aloud
• “e.g.” / “i.e.” → “for example” / “that is”
• Version numbers — “v2.4.1” will be read wrong by most engines; say “version two point four”

Specialised vocabularies are the other trap. Product names, people’s names and domain terms should be checked by listening, not by reading. If a term will appear in twenty decks, fix it once and note it in your style guide.

A review pass that catches almost everything

Read the script aloud, at conversational speed, before rendering. Not silently — aloud. Anything you stumble over will be worse when the voice hits it, because you can compensate for awkward phrasing and the engine cannot.

Then check three specific things:

1. Does anything assert more than the slide supports? A generated draft is inferred from the slide and can be confidently wrong. This is the one error that matters, because it ships a factual mistake to an audience that will believe it.
2. Are all proper nouns pronounced correctly? Listen to them.
3. Is any slide getting more words than it deserves? If a slide needs four minutes, it probably needs to be two slides — see our note on how long each slide should be.

The one habit that matters most

Write the way you would explain this slide to a colleague who is standing next to you and cannot see it. That is the register. Slide text is written for someone who can see it, which is precisely why reading it aloud sounds wrong.