I once rendered my first book in a single voice, and two chapters in I caught myself scrolling back up the page to figure out who had just spoken. Which was a bad sign, because I'd written the text. Since then I never push a dialogue-heavy book through render without a casting pass. Ten or fifteen minutes before render saves hours of "wait, who?" later.
What follows is what I've worked out for myself over time. Not rules — patterns.
When automation is fine, when it isn't
Most services these days auto-assign voices: parse characters out of the text, guess gender, age, type, match to descriptions.
Trust automation when:
- The text is short, under thirty thousand characters.
- There are no more than two to four speakers.
- It's non-fiction without dialogue.
- You're just running a draft to feel out the narrator voice.
Step in manually when:
- The book is long, and you'll be living with it for hours.
- There are characters you specifically care about (the main pairing in a fanfic, your protagonist).
- Automation visibly missed — gave a stern adult man a teenage voice.
- Two characters in dialogue end up sounding so similar you can't tell them apart.
Archetypes that work most of the time
I keep a few mental templates for casting. They're not "right," they just keep working as starting points.
The narrator. Neutral, mid-age, no accent, moderate pace. I default to male — it's the conventional choice in English audiobook tradition. For first-person novels with a female protagonist-narrator, female reads cleaner.
The young protagonist. Slightly higher than mid, lively, emotional. Not bassy. A seventeen-year-old shouldn't sound forty.
The female lead. Depends on character. "Strong, independent" — confident mid-range. "Sensitive, artistic" — softer, slightly below mid. Either way, no syrup.
The antagonist. A common mistake is reaching for an "evil voice." It doesn't land. Even, calm, cold, no obvious emotion is what works. Scarier than theatrical laughter.
The wise mentor. Older, usually male, slow pace, warm. The Gandalf template hasn't gone anywhere, and it still does the job.
Love interest. Mid-range, slightly more warmth than the narrator, no slipping into sweet. The line is thin and easy to cross.
Parents. Adult, mature voices. Mother — warm. Father — even, authoritative without weight.
Children. High, fast. Only when the text actually has children. A fifteen-year-old protagonist sounding eight is unintentionally funny.
The mistakes almost everyone makes
Too much contrast. Ten characters, all radically different voices, from squeaky teen to bass grandpa — your ears tap out twenty minutes in. Better: three or four anchor voices for the main cast and others built around them, with small timbre shifts. The book should sound like an ensemble, not a zoo.
Stereotyped picks. Villain: raspy and low. Fool: thin and high. Scientist: monotone. It works in cartoons; in audiobooks it goes from cute to caricature in an hour.
Ignoring the author's signals. If the text says "Anna's high voice" and you cast her low and velvety, complaints will follow. Before casting I scan the text for any explicit voice descriptions the author baked in. Five minutes, saves embarrassment.
Gender slips in auto-cast. Sometimes the automation routes a male role to a female voice or vice versa. A quick scan after auto-cast catches this in seconds.
Style hints
Most modern services let you describe the style on top of picking the voice itself. This works, and I lean on it.
What reliably lands: "cold, detached," "with a slight smirk," "slow, thoughtful," "fast, certain," "raspy, tired." Two or three words is enough. The longer the description, the worse the model handles it.
What doesn't land: long cinematic sentences like "the cold, aggressive baritone of an experienced officer who just received bad news." The model treats this worse than a plain "cold, firm." Less is more.
A separate note: don't overdose on emotion. Light coloring is good, especially when the scene is genuinely heavy. Whispers, sobs, full hysterics — almost never. You can listen to that for five minutes. You cannot listen to that across a fifteen-hour book.
Long books are their own pain
In 500-plus-page novels there's a real risk you'll start losing track of which secondary character has which voice. What works for me is a small table: character — voice — style hint — a couple of distinguishing notes. Something like: "Boris — Charon — 'tired, ironic' — always pauses before speaking."
For truly minor characters who appear in two or three scenes, I don't bother — let automation pick. Tuning "the bookstore clerk in chapter seven" isn't a good use of an evening.
For recurring characters, write them down. A hundred pages later you won't remember what you picked for the secretary in chapter four.
Listen to chapter one before pushing the rest
After the first chapter renders, listen all the way through. I check three things:
- Can I tell characters apart in dialogue without prompts? If three voices in a single scene blur, something needs to change.
- Do any two voices conflict? Sometimes my picks are so close they merge in a shared scene.
- Do characters sound the way I imagined? If I cast "cold antagonist" and hear theatrical menace, the style hint needs work.
Re-rendering is your friend
Decent services (we're one) let you re-render specific chapters or scenes. That's the safety net: in chapter five you realize Ira's voice isn't right — override — render only chapter five. The rest stays.
Faster, cheaper, and removes the fear of "now I have to redo everything." Don't be precious about pointwise fixes — they're a normal part of the process, not a sign you set things up wrong upfront.
The pre-render checklist
I run this myself before queueing a full book:
- Every key character got a manually picked voice.
- Voices are distinguishable in a single dialogue scene.
- Author-set voice descriptions in the text were respected.
- Style hints added for complex roles.
- The narrator doesn't pull attention to itself.
All five yes — push it. Most likely the result will hold up, and you'll only come back to this list for the next book.