How to Fix AI Audiobook Mispronunciations

A single wrong name can spoil an otherwise good audiobook, especially when it comes up two hundred times. AI voices are good at ordinary words but guess at names, invented words and words that change with meaning.

To fix an AI audiobook mispronunciation, tell the tool how the word should sound, using a respelling like “shi-VAWN” or a phonetic spelling in IPA, and apply it to every time the word appears. For words like “read” or “live”, which change with meaning, fix each one in place or reword the sentence.

This guide explains why AI voices mispronounce words, how to find every mistake, and how to fix them in ElevenLabs, Google Play, Amazon’s Virtual Voice and Audiobook Generator.

Why AI voices mispronounce words

AI voices learn pronunciation from huge amounts of text and speech. They’re very good at common words and guess at the rest. Mistakes cluster in a few places:

• Names. Siobhan, Nguyen, Saoirse, Llanelli: names from other languages often follow different spelling rules.
• Invented words. Fantasy and science fiction names have no right answer for the AI to find. ACX calls these “made up names, or invented languages that a narrator couldn’t look up”.
• Homographs. Words spelled the same but said differently: “read” (present or past), “lead” (to guide or the metal), “live” (to live or a live show), “minute” (time or tiny). Google lists “read”, “minute” and “present” as examples.
• Acronyms. An AI may say an acronym as a word when it should spell it out, or the other way round.
• Numbers and abbreviations. Years, fractions, Roman numerals, “Dr.” and “St.” Most of these are best fixed in the manuscript, as our guide to preparing your manuscript for AI narration explains.
• Accent. The same word can be right in one accent and wrong in another. Google’s example: “King Louis” sounds like “King Lewis” in American English but “King Louie” in British English.

Step 1: Find the words to fix

Don’t wait until you hear a mistake. Before you narrate:

1. List every name of a character, place or organisation in your book.
2. Add invented words, foreign phrases and technical terms.
3. Search for homographs that matter in your book, like “read”, “lead”, “live”, “tear”, “wind” and “bow”.
4. Note acronyms, and whether each is said as a word (NASA) or letter by letter (FBI).

Then listen. Play the first chapter with the text in front of you, and add anything else that sounds wrong to your list. Pay most attention to the first five minutes, because that’s the sample most stores play to shoppers.

Step 2: Write down how each word should sound

There are two ways to tell an AI voice how to say a word.

Respelling

Write the word the way it sounds, with the stressed syllable in capitals and hyphens between syllables:

Respelling works in every tool, and anyone can write it. It’s the method Amazon’s own help pages use: “ed-in-bruh” for Edinburgh, “led” for lead.

Phonetic alphabets (IPA and CMU Arpabet)

The International Phonetic Alphabet writes sounds exactly: “minute” as /ˈmɪnɪt/ (time) or /maɪˈnut/ (tiny). It’s more precise than respelling, but only some tools accept it, and only with some voices. CMU Arpabet is a similar system written in plain letters, which ElevenLabs recommends for some of its models.

If you don’t know IPA, look the word up in a dictionary such as Wiktionary, which gives IPA for most words and many names.

Our advice: start with respelling. Use IPA only when a respelling still doesn’t sound right and your tool supports it.

Step 3: Fix it in your tool

Audiobook Generator

Add the word to your pronunciation list with how it should sound, and every chapter uses it. A word can apply to this book or to every book you make, which suits your own name or a series’ places. You can also limit a word to one character, match it inside longer words, or match exact capitals only.

For a single wrong line, such as a homograph, read just that line again without redoing the chapter.

ElevenLabs

ElevenLabs uses pronunciation dictionaries, which you upload in Studio as a text or .PLS file. Each entry is either:

• An alias: a respelling, like “Claughton” said as “Cloffton”. This works with every model.
• A phoneme: the word in IPA or CMU Arpabet. This only works with some models, and its documentation disagrees on which. Multilingual v2, a common choice for audiobooks, doesn’t support phoneme tags, so use aliases with it.

Two more things to know: entries are case-sensitive, so add “Bree” and “bree” separately if both appear, and phoneme entries only work for single words. Our ElevenLabs audiobook guide shows the full workflow.

Google Play Books auto-narration

Google’s auto-narration editor has a pronunciation check that lists names, places, organisations and spelled-out words, and plays every place a word appears. To fix one, you can:

• Choose an alternative pronunciation from Google’s suggestions.
• Type a phonetic spelling.
• Enter it in IPA.
• Record yourself saying it, in some languages.

Amazon Virtual Voice

In Virtual Voice Studio, select the word, choose Edit pronunciation and type how it should be said, up to 200 characters. Tick Apply to all instances to fix every occurrence, including any you add later.

Apple Books digital narration

Apple doesn’t document a way for authors to edit pronunciation in its digital narration. The audio is made by Apple, so list names carefully when you nominate a book and check the result.

Step 4: Handle homographs one by one

Homographs are the one case where “fix it everywhere” doesn’t work, because “read” is right in one sentence and wrong in the next. Three ways to handle them:

1. Fix each instance. Most tools let you correct a single line. Use this for a few occurrences.
2. Reword the sentence. “She had read the letter” can become “She had already read the letter” if the AI keeps getting it wrong, or swap the word for a synonym.
3. Limit the fix. If a word is only ever said one way by one character, a fix limited to that character can work.

Step 5: Check the whole book

After fixing, listen again. Don’t just check the line you changed: a respelling can sound right in one sentence and odd in another. A good routine:

• Play several occurrences of each fixed name, in different sentences. Google’s editor can play them all for you.
• Listen to the first and last few minutes of every chapter.
• Ask someone who knows how the names should sound to listen to chapter one.

Common words to watch

FAQ

Why does AI mispronounce names?

AI voices learn how words sound from common text and speech. Unusual names, names from other languages and invented names follow rules the AI hasn’t seen, so it guesses.

How do I make an AI voice say a word correctly?

Add it to your tool’s pronunciation list or dictionary with a respelling, like “shi-VAWN” for Siobhan, and apply it to the whole book.

What’s the difference between an alias and a phoneme?

An alias replaces the word with a respelling the AI reads normally. A phoneme gives the exact sounds in a phonetic alphabet like IPA. Aliases work everywhere; phonemes only with some voices.

Can I fix “read” and “lead”?

Yes, but one at a time, because the right pronunciation depends on the sentence. Fix each line, or reword the sentence if the AI keeps getting it wrong.

Do I need to learn IPA?

No. Respelling works in every tool and fixes most words. Use IPA only for the few words a respelling can’t get right.

Ready to hear your book?

Make your list of tricky words, then hear how your book sounds. Upload your manuscript and narrate your first chapter free.