Generate a personal-finance chapter and you may hear "twenty nineteen" in one paragraph and "two thousand and nineteen" in the next. Both of those readings are correct. Side by side, they sound like two different editors worked on the book. Careful writers do this all the time: they clean the prose, leave the digits alone, and the digits are exactly where a speech engine has to guess.
Numbers are not decoration in audio. A wrong year, a wrong unit, or a skipped "not" next to a figure changes the claim. Text-to-speech will pick a reading. Your job is to pick it first, write it down, and use the same reading everywhere the same figure appears.
Find the strings that can move
Search the spoken script, not the outline. You want the characters a voice will meet.
- Years and dates. 2019, 9/11 as a date or as a fraction, March 3 vs 3 March.
- Quantities and units. 5 km, 5kg, 5 kg, "5 lbs." The space and the abbreviation change the guess.
- Money. $4.99, $4, 4.99 dollars, "a dollar."
- Ranges and ratios. 3-5, 3 to 5, 8/10, 8 out of 10.
- Identifiers that should stay digits or letters: versions, model numbers, error codes, section numbers.
- The small words glued to figures: not, only, about, less than, per.
The preflight scan in VoxAnvil flags numbers, abbreviations, currency, and dates, among other risks, and it suggests a spoken replacement. Use the flags as a net. Do not treat a suggestion as a decision. "Four point nine nine dollars" and "four dollars and ninety-nine cents" are both suggestions you might see in spirit. Only one of them matches how you talk about a price in that chapter. The pricing page of this product, for example, is written in dollars and cents. If you narrate those prices, say them the way the page states them, and do not round.
Decide a house style, then write it
Pick a house style for the project and put it on the series card or the pronunciation list. You can change the style. You cannot improvise it per paragraph.
A usable house style looks like this, as an illustrative example for a business book: years are "twenty nineteen," not "two thousand nineteen." Currency under a hundred dollars is dollars and cents. Percentages are the word "percent." Ranges are "to," not a dash. Units are words: "kilometers," not "k-m," unless you are quoting a label. Version numbers are digits in a row only if you have heard them and accepted them; otherwise write "version four."
Then replace or store. For a one-off lesson, write the words in the script: "about fourteen thousand cases." For a book, store the spoken form in Pronunciation Forge when the printed spelling has to stay. Forge is on Starter and Pro. Saved rules are substituted into the text before it goes to synthesis. Longer phrases apply first, so store "5 km" as a phrase if storing "km" alone would also touch "km" inside a code you did not mean to change.
If the printed book and the audio book can differ, prefer rewriting the spoken file. Audio does not need the digit "14,000" to look like a table. It needs the listener to hear a magnitude. "About fourteen thousand" is often better than a digit string the voice may group oddly. Keep the exact figure in the ebook.
The annoying cases
Dates. Decide month-day or day-month for the whole project and write the month name in the spoken file when a slash would be ambiguous. "3/4" is a date or a fraction. The voice should not have to guess, and neither should a listener in another country.
Ranges. A hyphen between numbers may be read as "minus" or skipped. Write "three to five days" if that is the claim. If you mean a score of three minus five, write that in words too. Do not leave both meanings on a hyphen.
Identifiers. An error code that must be typed later should be spoken in a way a person can type. Group the characters in the script: "A B C, one two three." A glob of letters is how codes get lost. If the code does not need to be typed, do not speak it. Put it on screen or in the notes. Podcast scripts especially benefit from this cut. The podcast script guide is the place for that split.
Units next to a decimal. "5.0" may be read as "five" or "five point oh." If the point matters, write "five point zero." If it does not, write "five" and stop showing precision you will not defend.
Negation. "Not 5 mg" is a different dose from "5 mg." Keep the "not" in its own short sentence if the voice keeps skating past it: "This is not five milligrams. It is one." Then listen. The listening pass is mandatory on dose, money, and law. A fluent voice can be fluently wrong.
Store, generate a sample, then commit
Do not generate the whole manuscript to test a number style. Build a twenty-second paragraph that contains one year, one price, one unit, and one negation. Generate it with the narrator you already chose. If the paragraph matches the house style, apply the same treatment to the book and generate chapters. If it does not, fix the style before you multiply the mistake by thirty chapters.
When you change a rule later, search for the old reading and regenerate those chapters. A new rule does not edit yesterday's audio file. Note the chapters in the log so a series does not publish two styles. The pronunciation list is the same habit applied to names. Numbers deserve their own section on that list so a name cleanup does not bury a dose.
This pass will not make a retailer accept your files. It will make the listener hear the figure you wrote. That is the standard. If a platform has its own rules about what you may narrate, those rules are on the platform. The studio's job is the spoken manuscript.
Frequently Asked Questions
Q: Should I spell out every number? A: Spell out the ones that are claims: years, money, doses, ranges, and anything with "not" attached. Leave a casual small count as a digit only after you have heard it. "Three steps" written as words is rarely worse.
Q: What about Roman numerals in chapter titles? A: If you want "chapter four," write "Chapter 4" or "Chapter Four" in the spoken file. Do not assume "Chapter IV" becomes a four. Hear it.
Q: Can the preflight scan choose my house style? A: It can suggest a spoken replacement for a flagged string. The house style is yours. Accept a suggestion only when it matches the style you wrote down.
Q: Will a better voice read digits more clearly? A: A clear voice helps, and it will still choose a reading. Two clear voices can say the same year two ways. The text decides. Audition after the numbers are stable, or you will blame a voice for a digit.
Pick one spoken form for each kind of number, write it down, and apply it to every chapter before you generate any of them. A year said two ways is a small error. A price heard one step off can be an expensive one.