Working with Audio
Turning a Script into a Voiceover
Synthetic speech has become good enough for narration, and what separates a usable result from an obviously robotic one is mostly how the script is written.
2 min read · Published 2 September 2026
Text to speech stopped sounding like a computer some years ago. What it still does is read exactly what you wrote, including the parts you would have adjusted while speaking, which is why the same tool produces natural narration from one script and something stilted from another.
Most of the improvement available is in the writing rather than in the settings.
Write it to be spoken
- Short sentences. A synthetic voice does not run out of breath, which is exactly why long sentences sound wrong: nothing forces a pause where a person would take one
- Punctuation is timing. A comma is a short pause and a full stop a longer one, so punctuation you would skip while writing has to be there
- Spell out what should be spoken. 1080p, GB, 24fps and similar are read inconsistently; writing them as words removes the guess
- Avoid words the reader has to see. Brackets, dashes and quotation marks either get read aloud or vanish, and neither is what you meant
Read the script out loud before generating anything. Every place you naturally paused and the text has no punctuation is a place the voice will not pause either.
Choosing a voice
Listen to a sample of the actual script rather than the demo sentence. Voices differ most on the material they were not trained to flatter: numbers, unusual names, and long clauses. A voice that sounds excellent reading a greeting can stumble through a technical paragraph.
Speed matters more than most people expect. Slightly slower than feels right when reading silently is usually correct for narration, because a listener cannot go back over a sentence the way a reader can.
Length and practicality
A long script is generated in pieces and joined, which is normal and inaudible when the pieces break at sentence ends. What it does mean is that a very long job takes real time, so it is worth generating a paragraph first and checking the voice before committing an hour of narration to it.
The voiceover tool here returns an MP3, which is what a video editor wants. If the audio needs to sit against picture, generating in sections that match your scenes is easier to work with than one long file, because a change to one section does not mean regenerating everything.
Where synthetic speech is the wrong answer
Anything where the voice is the point. Personal address, humour, emotion that has to land: these still need a person, and a synthetic reading of them is worse than a plain one.
It is also worth being straightforward with an audience. Narration that is clearly synthetic and not presented as a person is unremarkable; the same audio presented as a specific human being is not, and imitating an identifiable voice raises questions that have nothing to do with the technology.
Common questions
Why does the voice sound flat or robotic?
Usually the script rather than the voice. Long sentences with sparse punctuation give the reader nowhere to pause, so it delivers everything at one pace. Shortening sentences and adding the commas you would naturally pause at fixes most of it before any setting is touched.
How do I make it pause?
With punctuation. A comma gives a short pause, a full stop a longer one, and a paragraph break longer still. There is no separate pause control needed for narration; writing the punctuation you would speak is enough.
How should numbers and abbreviations be written?
As words, where the pronunciation matters. 1080p, fps, GB and similar are read inconsistently between voices, and spelling them out removes the uncertainty rather than hoping for the best.
Can I use synthetic narration on a monetised video?
Generally yes, provided the content is genuinely yours and adds something. What draws attention is mass produced narration over material with no original contribution, which is a content problem rather than a voice one. Imitating a real person's voice is a separate issue and best avoided entirely.