How to Fix AI Narration Pronunciation for Fantasy Audiobooks

Fantasy, sci-fi, and historical manuscripts break AI narration in one predictable place: invented and irregular proper nouns. Fix it in six passes:

1. Extract every proper noun and invented term from the manuscript.
2. Decide one canonical pronunciation per term.
3. Write it as a phonetic respelling your narration tool can act on.
4. Render a short test passage before committing the full book.
5. Transcribe the finished render and diff it against the manuscript to catch misses at scale.
6. Save the list as a series bible so book two matches book one.

You finished the manuscript. Narration is the last mile, and in genre fiction that mile has potholes: Caelith, Ysolde, Kh'raan, Beauchamp, Magdalen, Worcestershire. An AI narrator will pronounce every one of them confidently, and a meaningful share of the time it will be confidently wrong — and wrong the same way in all 340 places the name appears.

Human narrators solve this with a pronunciation list agreed with the author before recording. The AI workflow is the same list, just applied differently. Here's the process end to end.

Why invented names break AI narration

A text-to-speech system converts written text to sound in two stages: it works out how a word is pronounced, then produces audio for those sounds. For words in its lexicon, stage one is a lookup. For words that aren't — which is every name you invented — it falls back to guessing from spelling patterns learned across the language.

That guess is usually reasonable and occasionally absurd. Predictable failure classes:

• Apostrophes and hyphens in invented names. Kh'raan may get read as two words, or the apostrophe may be swallowed entirely.
• Non-English spellings inside English prose. A French or Welsh name gets English vowel rules applied to it.
• Real-world irregulars. Beauchamp (BEECH-um), Cholmondeley (CHUM-lee), Leicester (LES-ter). The spelling actively lies.
• Heteronyms. read, lead, bow, live, wound. Context usually resolves these, but not always in dialogue-heavy passages.
• Initialisms and units. AI read as a word, 2.5 km, Dr. at the start of a sentence.

The good news: this set is finite, it's knowable before you render anything, and fixing it is a one-time cost per series.

Step 1: Extract the term list from your manuscript

Don't work from memory. Pull the actual list.

The cheapest method is a capitalised-word frequency pass. In a terminal, from your manuscript as plain text:

12345

# Extract capitalised words not at sentence start, sorted by frequency
grep -oE '[^.!?]\s+[A-Z][a-zA-Z'"'"'-]{2,}' manuscript.txt \
  | grep -oE "[A-Z][a-zA-Z'-]{2,}" \
  | sort | uniq -c | sort -rn | head -100

Expected output: a ranked list — character names at the top with counts in the hundreds, place names in the dozens, one-off terms in the tail.

123456

 412 Caelith
 287 Ysolde
 143 Vareth
  61 Kh
  12 Aunt Beauchamp

That 61 Kh is exactly the signal you want: the apostrophe split Kh'raan in two. Note it.

Work the top 40–60 terms. Anything appearing three times or fewer you can handle inline later. If your manuscript lives in .docx or .epub, export to .txt first — most word processors do this from Save As.

Step 2: Choose one canonical pronunciation per term

This is an authorial decision, not a technical one, and it's the step writers skip. Say each name out loud and write down how you actually say it. Two rules:

• One pronunciation per term, for the life of the series. Readers notice drift between books more than they notice an unusual choice.
• Write it down even when it's obvious. In eighteen months you'll be drafting book three and you will not remember whether Vareth rhymed with "Gareth" or "breath."

If a name has an in-world pronunciation that differs from the obvious English read, that is a fact about your world and belongs in the bible next to eye colour.

Step 3: Write respellings that actually work

You have two notation options. The International Phonetic Alphabet is precise but unforgiving — one wrong symbol and you get worse output than you started with. Plain phonetic respelling is more robust for most authors: write the name as an English speaker would spell it if they'd only ever heard it.

Two habits worth forming. Use hyphens for syllable boundaries and capitals for the stressed syllable — ee-ZOLD gets you both shape and emphasis in one token. And avoid respellings that collide with real English words unless the collision is exactly right: respelling Vareth as Vah-reth is safer than Var-eth, which can pull toward a different vowel.

Step 4: Apply the respellings in Audiobook Studio

In Audiobook Studio, work on a copy of your manuscript — never your master file. Find-and-replace each term with its respelling across the working copy, then import.

A few practical notes:

• Match case-sensitively. Read at a sentence start and read mid-sentence may need different handling.
• Replace in dialogue and narration both. It's easy to fix a name in narration and miss it inside quoted speech.
• Keep a mapping file. A two-column CSV (original,respelling) is your bible and your undo. It's also what you'll reuse in book two.

123456

Caelith,Kay-lith
Ysolde,ee-ZOLD
Kh'raan,Krahn
Beauchamp,Beecham
Vareth,Vah-reth

The same approach works for shorter pieces in text to speech — a chapter sample, a podcast intro, a book trailer script.

Step 5: Render a test passage before the full book

Never render twelve hours to find out an apostrophe broke. Pick a 300–500 word passage that is dense with your problem terms — a scene where three named characters are in a named place — and render that alone.

Expected output: roughly 2–3 minutes of audio. Listen with the manuscript open and mark every term as correct, wrong, or borderline.

Most authors need two or three test rounds. That's normal and it's cheap: a few minutes of audio per round, against a full book render you'd otherwise have to redo.

On the free tier — 1,000 credits a month, no card — test passages are exactly the kind of work the allowance is sized for. Do your pronunciation loop free, then render the book on a paid plan once the list is locked.

How do I catch mispronunciations across a twelve-hour book?

You don't do it by listening. You do it by diffing.

Run the finished render back through speech to text to get a transcript of what the narrator actually said, then compare that transcript against your manuscript. Where the narrator mangled a name, the transcript shows the mangling — Kh raan as two words, Bow-champ instead of Beecham — and it shows it with a timestamp.

12345

# Normalise both sides, then diff
tr '[:upper:]' '[:lower:]' < manuscript.txt | tr -s ' \n' '\n' | sort -u > a.txt
tr '[:upper:]' '[:lower:]' < transcript.txt | tr -s ' \n' '\n' | sort -u > b.txt
comm -23 a.txt b.txt | head -50

Expected output: words present in your manuscript but absent from the transcript. Most are transcription noise. The ones that matter are your proper nouns — a name in that list means the narrator said something else, and you now know which chapter to check.

This will not be perfect; transcription has its own error rate, and homophone-level errors slip through. But it turns "listen to twelve hours and hope" into "check nine timestamps." That's the difference between a QA pass you'll actually do and one you'll skip.

Does ACX reject an audiobook for mispronunciations?

ACX's technical requirements are measurable ones: peak level, noise floor, RMS range, room tone, file format, per-chapter structure. A mispronounced invented name is not a technical failure and will generally pass automated and human QA, because nobody at the retailer knows how your fictional city is meant to sound.

Which means the enforcement is reviews. "The narrator says the main character's name wrong" is a recurring complaint in genre audiobook reviews, and it's the kind of one-star that doesn't decay. Treat pronunciation as a reader-experience gate, not a compliance one — the retailer won't catch it, and that's precisely the problem.

The technical side is separate and worth handling on its own terms: mastering targets, chapter splitting, opening and closing credits, and the retail sample.

Common errors and fixes

Step 7: Turn the list into a series bible

Keep the CSV with the manuscript, in version control or the project folder — not in your inbox. Add to it every time a new name enters the world. When book two starts, the mapping applies on day one and the narration matches book one without anyone having to re-derive it.

This is the part that compounds. The first book costs you an afternoon of pronunciation work. Every book after that costs a find-and-replace.

FAQ

Do I need IPA, or is plain respelling enough? Plain respelling is enough for the large majority of names. IPA is worth the effort when you need a sound English spelling can't represent — a rolled R, a non-English vowel — and you already read IPA fluently. If you don't, a wrong IPA symbol produces worse results than a plain respelling.

How do other AI voice tools handle this? Several offer some form of pronunciation dictionary or lexicon — ElevenLabs, Murf, and Play.ht have all documented features in this area; check each vendor's current docs, since the specifics change. The underlying job is identical everywhere: decide a canonical pronunciation, express it in a notation the engine acts on, and verify with a test render. The list you build is portable even if you switch tools.

Will fixing pronunciation change the voice or performance? No. Respelling changes what sounds get produced, not the voice, pacing, or delivery. Your narrator sounds the same — it just says the name right. You can hear how different voices handle your terms in AI Narrator before you commit to one.

What if a character's name is deliberately pronounced two ways? Use two distinct respellings and apply each in the right scenes. If one character says it wrong on purpose, that's a per-scene replacement, not a global one — do it after the global pass so you don't overwrite it.

How long does this take for a 90,000-word novel? For a first book, budget two to four hours end to end: an hour on extraction and decisions, an hour across two or three test rounds, and the rest on the transcript diff and spot checks. Later books in the series are typically under thirty minutes.

Does this matter for non-fiction? Less, but not zero. Non-fiction has its own set: author surnames, technical terms, place names, acronyms, and citations. The workflow is the same, the list is usually shorter.

Start with a test passage

Pull your top forty proper nouns, write the respellings, and render one dense scene. If it comes back clean, render the book.

The free tier gives you 1,000 credits a month with no card required — enough to run the whole pronunciation loop before you spend anything. Open Audiobook Studio, import a chapter, and check the names.