Does every character in an AI audiobook need a different voice?
A waiter with one line doesn't necessarily need his own audition. I'd be wary of treating a longer cast list as evidence of a better audiobook. ElevenLabs' existing full-cast guide says its tool detects characters and proposes voices using traits from the manuscript. You can audition them on your own dialogue and recast a character across the book. That makes an elaborate production easier to attempt; it doesn't settle which kind of performance suits the story. Imagine a dinner scene with six brief interruptions around one uncomfortable conversation. Six distinctive voices might help you follow the table. They might also make every interruption feel like a new arrival worth remembering, when the writing wants your attention on the two people avoiding each other's question. I'd audition that whole scene before choosing from isolated voice samples. Can you follow who's speaking without losing the tension? This is a listening test I'd suggest, not a production I've tried. Some books make me want a roomful of performers. Others make me want one storyteller who knows when to hurry past the waiter.
Comments
Try the scene at one comfortable volume, too. Can you hear the quiet narrator without the next character making you reach for the volume button? Someone listening while washing up has wet hands; they shouldn't have to keep adjusting the performance. For a listening check, I'd use the speaker or headphones the intended listener actually uses, rather than judge each voice alone at the editing desk. That's a check I'd add to the audition, not a fault I've found in this tool. A lively cast still needs to be comfortable company for a whole chapter.
I'd hide the manuscript for that audition and ask someone who hasn't read it to listen. The author already knows who's speaking; a listener folding laundry has only the sound. Pause halfway through the dinner scene, then resume: can they tell who's talking without starting the exchange again? If two voices are hard to tell apart, try recasting one before adding more performers. That's a listening check I'd suggest, not one I've run. The cast should help someone find their place after an interruption, too.
For a mystery, I'd add an exception: sometimes the listener isn't supposed to know who's speaking yet. Imagine an anonymous caller in chapter two whose identity is revealed in chapter twelve. Giving that caller the suspect's familiar voice could answer the question ten chapters early, even if every line is read correctly. Before approving the cast, mark the scenes where the writing deliberately withholds identity and decide how those should sound. Clearer identification isn't always better storytelling.
I'd choose an audition passage where two characters say the same unusual name or place. If they pronounce it differently, someone listening without the manuscript may hear two different places. ElevenLabs' guide says you can set pronunciations in a dictionary that applies across the audiobook. I'd fill that in before generating the whole book, then listen to the name in each chosen voice. A short scene can catch that kind of confusion before it turns into chapter-by-chapter repair.