#ad #ElevenAgentsPartner — this post contains affiliate links
Disclosure: I'm an ElevenLabs affiliate partner. Links in this post go to articles on my own site, infinityme.dev, and some of those pages contain affiliate links. If you buy through them I may earn a commission at no extra cost to you.
A ten-minute YouTube voiceover is a sprint. A novel is a marathon: eight to twelve hours of audio where every small quirk repeats until it's all the listener can hear. A name pronounced two different ways. A narrator who sounds a little brighter in chapter 14 than in chapter 3. Dialogue read in the same flat tone as the weather description.
That's why I don't judge narration voices on demos anymore. I treat an audiobook like any other long-running build: prepare the inputs, lock the config, run it in small chunks, and test as I go. Here's the checklist I use.
What actually matters over eight hours
A voice that wins a 20-second sample can lose over a full chapter. These are the things I listen for:
- Consistency. Same pitch, energy and pace in chapter 20 as in chapter 1. Some engines drift between generations, and you only notice when chapters play back to back.
- Pacing. Short pauses at commas, longer ones between paragraphs, no rushing through long sentences.
- Dialogue. Fiction lives on it. Quoted speech and questions should shift a little instead of sounding like narration.
- Pronunciation control. Character names and invented places have to come out the same way every time, and you want to fix them once, not sentence by sentence.
- Paragraph-level editing. Being able to regenerate one paragraph without re-rendering a whole chapter saves hours.
- A clear commercial license, if you plan to sell the book.
For most indie authors I'd start testing with ElevenLabs. Its voices handle emotional range well, which matters most for fiction, and its Studio editor is built for long projects: you organize a book into chapters and regenerate single paragraphs. The site's pricing notes, checked on the ElevenLabs pricing page in October 2026, list 3 Studio projects on Free and 20 on Starter. For calm, even non-fiction (business books, course companions) there are other tools with steady studio voices that are worth auditioning too. My best AI voice for audiobooks guide goes through the options.
Step zero: treat the manuscript as input data
Most of the bad narration I've heard wasn't the voice's fault. It was the text. Page numbers read aloud. Footnote markers turned into "bracket twelve". A character name the model guessed at differently in two chapters.
So before generating anything, I run each chapter (one .txt file per chapter) through a small script. It does three things:
- Strips obvious junk: footnote markers like
[1]and lines that contain only a page number. - Counts characters per chapter, because TTS is billed per character, spaces and punctuation included.
- Lists capitalised words that show up mid-sentence and never in lowercase. Those are usually names and places, which are exactly the words you need to test first.
import re
import sys
from collections import Counter
from pathlib import Path
RETAKE = 1.15 # budget ~15% extra for regenerated paragraphs
def clean(text):
text = re.sub(r"\[\d+\]", "", text) # footnote markers like [1]
text = re.sub(r"^\s*\d+\s*$", "", text, flags=re.M) # lone page numbers
text = re.sub(r"\n{3,}", "\n\n", text)
return re.sub(r"[ \t]+", " ", text).strip()
def name_candidates(text):
"""Capitalised words seen mid-sentence that never appear in lowercase."""
lower = set(re.findall(r"\b[a-z]+\b", text))
found = Counter()
for m in re.finditer(r"\b[A-Z][a-z]{2,}\b", text):
before = text[:m.start()].rstrip(" \"'\u201c")
if not before or before[-1] in ".!?:\n":
continue # sentence or line start: probably just grammar
if m.group().lower() not in lower:
found[m.group()] += 1
return found
total, names = 0, Counter()
for path in sorted(Path(sys.argv[1]).glob("*.txt")):
text = clean(path.read_text(encoding="utf-8"))
total += len(text)
names += name_candidates(text)
print(f"{path.name:<16}{len(text):>9,} chars")
print(f"{'TOTAL':<16}{total:>9,} chars ~{round(total * RETAKE):,} credits with retakes")
print("Check pronunciation:", ", ".join(w for w, _ in names.most_common(10)))
Here's the real output on two short test chapters I wrote for this post:
01-arrival.txt 378 chars
02-archive.txt 319 chars
TOTAL 697 chars ~802 credits with retakes
Check pronunciation: Vandermeer, Okonkwo, One, Teodora, Two, Kessrin
It's a heuristic, and it shows. "One" and "Two" come from the "Chapter One" headings. And it missed Ilse, the main character, because in my sample she only ever appears at the start of a sentence. So the list is a starting point you read through, not a final answer. Still, it caught every invented place name, and those are the ones a voice model is most likely to guess wrong.
Turning characters into a budget
ElevenLabs bills text to speech at about 1 credit per character on standard models, per their pricing page. A word averages roughly 6 characters including the space, so a 70,000-word novel is around 420,000 characters, or roughly 480,000 credits once you add retakes.
The plan figures below are what my site lists, checked on the official ElevenLabs pricing page on October 21, 2026. Check the current page before you buy:
- Creator: $22/month ($11 first month), 121,000 credits. Unused credits roll over for up to two months while you stay subscribed, so you can narrate a novel slowly, chapter by chapter, over a few months.
- Pro: $99/month, 600,000 credits. One month covers a typical novel with room for retakes if you want it done fast.
Professional Voice Cloning starts at Creator, which is the main reason to pick it if you want your own voice on the book. If you'd rather not do this math by hand, there's an ElevenLabs cost calculator on my site that does it in the browser.
Licensing: the trap that costs you a re-render
Two separate rulebooks apply.
Your voice tool's license. On ElevenLabs, paid plans include a commercial license and the Free plan doesn't. The catch that bites people: content generated on Free stays non-commercial even if you upgrade later. The license follows the plan you generated on, not the plan you have on launch day. Use Free to audition voices and test your name list, then generate the real narration on a paid plan. The details are in ElevenLabs free vs paid.
Each retailer's rules. Audiobook stores set their own policies on AI narration, and they've changed more than once. Before spending credits, read the current guidelines for every store you plan to use. Look for whether AI narration is accepted at all, whether you have to label it, the technical specs (format, bitrate, loudness, room tone at the start and end of each file) and whether a platform's own AI program limits selling elsewhere.
The workflow, like a build pipeline
- Clean the manuscript with the script above. Spell out abbreviations the way they should be spoken.
- Test every name from the list, plus the ones it missed. Fix spellings or phonetic hints now, not in chapter 18.
- Audition on real text. Generate a full page of dialogue and a full page of description. Listen on headphones and on a phone speaker.
- Lock your settings and write them down, like pinning a dependency version. For long-form, slightly higher stability usually helps consistency. I explain what the sliders do in ElevenLabs voice settings.
- Generate chapter by chapter and listen to each before moving on.
- Fix at paragraph level. Don't re-render a chapter, and don't switch voices because one paragraph sounded off.
- Master and export each chapter to your retailer's loudness and room-tone specs.
- Do a final full listen, ideally with someone else's ears.
Mistakes I'd avoid
- Producing on Free and planning to "upgrade before publishing."
- Changing voice or settings halfway through the book.
- Skipping the pronunciation pass.
- Leaving retailer specs until the end and re-exporting everything.
- Not budgeting 10–20% extra for retakes.
AI or a human narrator?
AI narration makes the most sense when budget is the real barrier, for non-fiction and backlist titles that would never get an audio edition otherwise, and for testing whether a book has an audio audience at all. A good human narrator still brings interpretation and character work that AI can't fully match, especially in dialogue-heavy fiction. Plenty of authors use AI for the backlist and save human narration for their main titles. I wrote more about that trade-off in AI voice vs hiring a voice actor.
The full guide, with a per-book credit table and the step-by-step for Studio, is here: Best AI voice for audiobooks.
Top comments (0)
Some comments may only be visible to logged-in visitors. Sign in to view all comments.