Writing for the Ear: Turning a Page into Audio

AI voices crossed the line a while ago. They are no longer the reason your audio sounds wrong. The script is.

Text written for the eye assumes things audio cannot give: headings that show structure, bullets that group ideas, and the ability to glance back at a sentence you missed. A listener has none of that. They get one pass, in order, at your pace.

One idea per sentence

Written prose can carry a subordinate clause, a parenthesis and a qualifier and still be followed, because the reader can slow down. A listener cannot. By the time the sentence ends they have lost where it began.

Split them. Shorter than feels natural on the page is usually right in the ear.

Say the structure out loud

You cannot speak a bullet point. What you can do is signpost:

"There are three things worth knowing here. The first is..."

That does the job a list does on the page, and it also gives someone who drifted a way back in. Headings become spoken transitions: "So that is the setup. Now, what to actually do about it."

Spell out anything hard to hear

Numbers, acronyms, URLs. "£1,250" read by a synthesiser can come out as digits. "ROI" may be spelled out or pronounced as a word depending on the engine.

Write what you want heard: "one thousand two hundred and fifty pounds", "return on investment", "our website, slash pricing".

Repeat more than feels comfortable

On the page, repetition is a flaw. In audio it is a courtesy. A listener washing up will miss things, and a brief restatement of the key point before you move on costs a sentence and saves them rewinding.

Read it aloud yourself first

This is the whole quality-control process and it takes three minutes.

Read your script out loud. Wherever you stumble, run out of breath, or get bored — the script is wrong there, and no voice model will rescue it. Fix those spots before you generate anything.

Where AI audio works well

Turning written material you already have into a listenable version. Consistent narration across a series. Making long documents accessible. Drafting so you can hear the shape before recording it yourself.

Where it does not: anything that needs genuine warmth or a relationship. Synthetic voices are good at clarity and poor at intimacy. If the point is you, record yourself.

Say that it is AI

If a listener discovers a voice they thought was human was not, you lose more trust than the production value ever earned you. A short line at the start costs nothing.