Fixing a mispronounced name in one chapter fixes it in one chapter. That is the trap in a long fantasy manuscript: the hero's city goes out to early listeners said three different ways, you correct the spelling in chapter one and regenerate, and then you type the old spelling again in chapter four. The fix stayed right where it was made. A pronunciation list would have carried it forward.
A pronunciation list is a decision record. For each tricky string, you write what the listener should hear. You store that pair where the next generation will use it. You do not rely on memory, and you do not rely on the voice to "get it" the second time. Voices do not learn your book between chapters.
What goes on the list
Start with words that are expensive when they are wrong.
- People, places, and invented terms. If you built the word, the engine did not grow up with it.
- Acronyms. Decide whether you want letters ("N-A-S-A" is usually the word "NASA," which is its own decision) or a phrase. "US" can be "U-S" or "United States." Pick one for the project.
- Words that are also common words. A character named "Read" may need "Reed." A town named "Polish" may need a different stress than the nationality.
- Phrases that include a shorter word you might also replace. If you replace "AI" inside "Aiden," you will damage the name. Phrase length matters, and the studio already applies longer saved phrases first so a short rule does not cut the middle out of a longer name. You should still avoid sloppy short rules.
Do not put the whole dictionary on the list. A rule for "the" or "a" will haunt you. Add a word when you have heard it fail or when you know it is unusual. Narration QA on Pro mentions acronyms that have no saved rule. That list is a prompt, not an order to expand every capital letter. "OK" may already be fine. "SQL" may not be, if you want "sequel" rather than three letters.
Decide the spoken form in writing
Write the spoken form as plain letters a stranger could read. "Kee-AH-ra" is a decision. "Say it nicely" is not. You do not need a linguistics degree. You need a form you can check by ear.
For an illustrative example, picture a memoir that mentions a grandmother from "Gloucester" and a street spelled "Marlborough." You decide "GLOS-ter" and "MARL-bruh," write both next to the spellings, and store them. When the voice says the spelling instead, you can point at the rule. When the voice says your form and you hate it, you change the rule and regenerate the lines that contain it. The list is allowed to be revised. It is not allowed to be vague.
If two characters should say the same place differently, you have a harder problem. A single project rule applies to the phrase wherever it appears. You cannot store two spoken forms for one spelling and expect the studio to know who is talking. In that case, change the text in the line that needs the unusual reading, and leave the rule for the common reading. Be honest in the list: "rule is the narrator's reading; dialogue on page 40 is spelled out."
Where the list lives in VoxAnvil
The preflight scan reviews manuscript text and returns flags for numbers, abbreviations, contractions, names, punctuation, currency, dates, formatting, and other spots a speech engine is likely to misread. Each flag includes the original phrase and a suggested spoken replacement. That suggestion is a starting point. Accept it only if you would say it.
Pronunciation Forge, on Starter and Pro, stores the reading you actually want. When a chapter is generated, saved rules are substituted into the text before it is sent for synthesis. The free plan can still generate a short amount of audio, and you can type the spoken form directly into the manuscript if you do not have the dictionary yet. The dictionary is what keeps chapter nine from undoing chapter one.
Run the scan, store the rules you mean, then generate. Generating first and collecting failures is how you pay for the same chapter twice. One chapter credit is typically up to 2,500 words. A rule added after a render does not change the file you already downloaded. Regenerate the chapter, or the section, that contains the phrase.
There is a longer walkthrough of the scan-then-store order on the preflight method page. This article is the habit: the list exists before the render.
Keep the list with the series
Copy names forward when a series continues. Do not copy rules you do not recognize. Read them. A rule that replaces "Dr" with "doctor" is right in a hospital book and wrong in a book where "Dr" is the start of a filename. Rules travel badly when they are treated as a blob.
Keep a column for "heard on" and the date you checked. A rule you have not heard is a guess. The listening pass is how a guess becomes a confirmed line. If you change a rule, note the chapter you regenerated so you do not leave episode three on the old reading and episode four on the new one.
Numbers and units are a special part of the same list. Do not bury them in a pile of names. Give them their own lines, using the decisions in the numbers guide. A date said as digits in one chapter and as a month name in the next feels like two books.
Frequently Asked Questions
Q: Can I write pronunciations in the international phonetic alphabet? A: You can put those characters in a spoken-form field. Use them only if you have checked that the voice says what you intend. A respelling you can read aloud is easier to audit. The test is the audio, not the alphabet.
Q: Do rules apply inside words? A: Short rules can damage longer words if you are careless. The studio applies longer phrases first and uses word boundaries when the phrase starts and ends with a word character. Still, do not store a two-letter rule unless you have searched the manuscript for those letters.
Q: Will the preflight scan fix the book automatically? A: It flags and suggests. You choose what to store. It does not approve a retailer file, and it does not replace listening.
Q: What if the voice ignores a clear respelling? A: Try a more phonetic spelling, generate that one sentence, and listen. If it still fails, change the sentence so the spelling is the pronunciation, or choose a voice that can say the simpler form. Document what you tried so you do not repeat it.
Start the list with the names you have only ever seen written down. Those are the ones a voice will guess at, and listeners remember the guesses.