Narrating a Kids' Storybook With ElevenLabs (And Getting the Voices Right)
Three characters, one narrator, a four-year-old as the quality inspector. The voice choices that made AI narration pass the bedtime test.
My test audience for this project was the harshest available: a four-year-old who accepts zero robotic voices and interrupts on demand. ElevenLabs narrated her storybook, and getting it bedtime-ready taught me most of what matters about character voicing.
The structure that works: one narrator voice carrying the prose, one distinct voice per recurring character, nothing more. I picked the narrator for warmth and a slightly slower default pace, then chose character voices that differ on two axes (pitch and texture), not one. Two characters who differ only in pitch sound like the same person on cheap speakers, which is exactly the failure my first draft had.
The text needs directing, not just pasting. Dialogue tags become performance instructions: 'she whispered' belongs in the delivery, so restructure to put the whisper in the sentence the voice actually says. Pause punctuation matters more than you'd think; em-dashes and ellipses at story-beat moments create the dramatic beats that keep a small listener tracking.
Kids' content has a special quality bar: they re-listen. The same chapter forty times will expose every synthetic artifact a single listen hides. My fix was regenerating any line that made her frown, which sounds unscientific but converged fast. About one line in ten needed a second take, usually for pacing rather than voice quality.
The bedtime verdict, after two weeks: requested nightly, preferred over my own narration, which stings precisely. The workflow is an evening of setup and then chapters on demand. For family projects, prototype content, or any story where hiring a narrator is absurd, this is now the obvious first move.