Recording, Editing and Voice

The AI wins in production are editorial rather than generative — and they are large. For talking-head content, transcript-based editing routinely halves edit time.

Record in one pass

Do not stop for mistakes. Pause, repeat the sentence cleanly, and keep going. Transcript-based editing makes removing the flubbed take trivial, so stopping only breaks your rhythm.

This single habit changes recording from stressful to routine.

Edit the transcript, not the timeline

• Delete a sentence in the transcript and it disappears from the video
• Remove filler words in bulk rather than one at a time
• Apply noise reduction and voice enhancement in one pass
• Find the section you want by searching the words rather than scrubbing

Where synthetic voice is genuinely good enough

• Localisation — a version in a language you do not speak, where the alternative is that version not existing
• Corrections — fixing one wrong word without a re-record session
• Training and explainer narration, where the register is neutral
• Temp tracks for timing an edit before the real recording

Where it is not

Anything emotional, and anything personality-led. If people follow you for you, a synthetic version of your voice reads as a shortcut — and audiences respond badly when they discover it.

Over long durations, small unnatural rhythms accumulate into listener fatigue even when no individual sentence sounds wrong.

The consent rule

Clone your own voice. Get explicit written permission for anyone else's — colleagues, clients, public figures, all of them. There is no version of this where 'it was only a test' is an adequate answer.

If a listener would reasonably assume a human spoke, and knowing otherwise would change their trust, say so.