Writing for the ear, not the eye

A text that is perfect to read can be painful to listen to. Out loud, you can't reread a sentence or look at a table: the text has to carry itself. That's the main reason so much generated audio sounds "robotic".

The rules that change everything

• Short sentences, and a rhythm that varies: one long, two short. "Twenty-six percent. In February, it was less than one percent!"
• Numbers written out when they matter: the voice no longer stumbles, and you control exactly what is said.
• Punctuation as breathing: ellipses create suspense, exclamation marks add energy. Synthetic voices use them.
• No web addresses, no bullet points, no tables: say "every article is on alyriaz dot com".
• Talk to someone: questions to the audience, reactions ("And honestly…"), natural transitions ("Okay. Let's take a breath…").

Who is speaking?

Pick an editorial voice and stick to it. At Alyriaz, the podcast speaks on behalf of the outlet: "we", "us", never "I". A script that mixes both loses its listener.

Checking without listening ten times

A production tip: have the resulting audio transcribed by a speech-recognition tool, then compare it with your script. A mispronounced word jumps out. On the Alyriaz podcast, that's how we caught "label" being heard as "à la belle": we replaced the word with "mark of quality".