Prepare a Manuscript for AI Narration: Requirements First, Then a Cleanup List

Before you generate any audio, get the source file right: an .epub for Google (whose create page says it must be live on Google Play), a reflowable EPUB with a navigation table of contents for KDP Virtual Voice, and an EPUB for Spotify's tools. Then clean the text so a synthetic voice has nothing to stumble on. The first part below is documented requirement. The second is practical advice about what to check, with no measured error rates behind it.

The file requirements each route documents

As of 2026-11-08, the stores that name a source format name EPUB, and some add conditions. This section is the requirement; later sections are advice.

Each store page has the full rule: Google Play auto-narration, KDP Virtual Voice, Spotify's creation tools and Apple Books.

Because the Google and KDP routes can read your live ebook (KDP requires a live eBook; Google's create page requires a live .epub, while its program policies also allow a newly uploaded one), a fix you make for narration may change the ebook itself. Decide whether you can accept that, or whether you will keep a separate narration copy for tools that take an uploaded file.

Decide where each edit will live

Choose where each edit goes before you start, because Google and KDP work from a live ebook and Spotify will not take a replacement EPUB after publishing. As read on 2026-11-08:

Whichever route you take, sort each change into one of two piles:

• Fixes that help both formats. Typos, inconsistent names, missing end punctuation, real chapter headings. Make these in your master file.
• Fixes for the ear only. Respellings, "Doctor" for "Dr.", removed figure references. Keep these in a narration copy or in the tool, so a phonetic "shi-VAWN" never reaches a reader's ebook.

Structure: what to check in the EPUB

Check the EPUB's structure first, because KDP lists a table of contents as an eligibility item and every route depends on chapters splitting where you expect.

1. Table of contents. Confirm it exists and links every chapter. KDP names NCX or Kindle Interactive; KDP lists it as an eligibility item.
2. One heading per chapter. Check that chapter titles are real headings, not styled paragraphs, so chapters split where you expect.
3. Front and back matter. Decide what is read aloud. Copyright pages, dedications, "also by" lists and acknowledgements are your call. Remove what a listener should not hear.
4. Chapter and paragraph size. Compare against the Studio limits above if you use ElevenLabs. A very long paragraph is the likeliest to hit the 5,000-character cap. At an assumed six characters per word including spaces, that cap is about 830 words (5,000 ÷ 6), which is our estimate.
5. Forewords and quoted passages. Google's editorial tips say to give the writer's name at the start of forewords, afterwords and passages. On the page, a byline does that job. In audio, the listener hears nothing unless the text says it.
6. Anything excluded automatically. Google excludes footnotes, footnote references, tables, figure text and page breaks by default, and you can add them back. If a footnote carries something the listener needs, move it into the sentence before you generate.

Pre-flight bench

Check a passage before you spend credits

Paste a chapter or a tricky passage. The bench flags the things this guide says to check, and can read it aloud in your device's built-in voice so you can catch names and numbers that come out wrong.

Text that does not read aloud well

These are checks, labelled as reasoning. Each item is something a human reader would silently skip or interpret.

• URLs and email addresses. Decide how each should be spoken, or cut them. A web address read letter by letter is rarely what you want.
• Footnotes and endnotes. Decide whether to read them, move them, or drop them. Superscript markers can end up in the audio as stray numbers.
• Tables, charts and image captions. These do not translate to audio. Rewrite essential data as a sentence. Apple's page and KDP's eligibility list both flag image or table heavy titles, which is covered in nonfiction versus fiction.
• References to the page. "See the chart above", "as shown overleaf" and "turn to page 40" mean nothing in audio. Google's tips say to edit or remove references to charts, diagrams, tables and long lists.
• Symbols and abbreviations. "&", "%", "No.", "St." and "Dr." can each be read more than one way. Spell out the ones that matter, and check how a name like "Barnes & Noble" comes out.
• Numbers, dates and units. Decide whether "1984" is a year or a quantity, and whether "3/4" is a fraction or a date. Write what you want spoken. ElevenLabs' best-practices guide lists phone numbers, currencies, dates, times, addresses, URLs and unit abbreviations as items its models can struggle with.
• Roman numerals and ordinals. Chapter "IV" and a king "Henry VIII" are read differently. Check both.
• Scene breaks. Symbols such as "***" may be read aloud or ignored. Replace them with a consistent break your tool handles, then listen for it.
• All-caps words and stylised text. Capitalised shouting, drop caps and spaced-out letters may be mispronounced or spelled out.
• Foreign words, names and invented terms. Build a list now. The fixes for them are in pronunciation and pacing. ElevenLabs' audiobooks page says pronunciation entries are pre-filled from the manuscript, which helps only if the names are consistent.

Before and after: sample edits for a narration copy

The examples below show the kind of change to make in a narration copy. We wrote them for this page, the names and figures are invented, and none is a measured result. How a tool reads the "before" text depends on the tool and even the model: ElevenLabs' guide says "$1,000,000" comes out as "one million dollars" on Multilingual v2 and "one thousand thousand dollars" on Flash v2.5.

Two cautions. Do not expand every number on reflex, because a tool may already read common forms well and a spelled-out figure is harder to proof against your print edition. And capitals are not always a problem: ElevenLabs says capitalization increases emphasis on its v4 model, so "NOT" might be doing what you want there. Try a paragraph first.

A search list for finding these in your file

Your word processor's find function will catch most of this list. Work through the narration copy once per term, and decide each case rather than using replace-all, since "St." can be Street or Saint.

ElevenLabs' guide also describes normalizing text with an LLM prompt or regular expressions, aimed at developers. If you run any automated pass, compare the result against your original word by word, because a model can change more than the numbers.

Punctuation and sentence shape

Punctuation drives phrasing in most text-to-speech, but how each tool treats commas, dashes and ellipses is something to check by listening, not assume. Fix obvious problems at the source: missing end punctuation, run-on sentences, and dialogue with no tags. Splitting long sentences is covered in the pacing page linked above.

ElevenLabs is one vendor that documents this: its guide says dashes and ellipses can stand in for short pauses but are less consistent than a break tag, and that its v3 and v4 models do not support break tags at all. An illustrative split, written by us:

• Before: The clinic, which had been closed since the flood that took the east wing in the spring of that year, reopened with half its staff, no records and a roof that leaked whenever the wind came off the river.
• After: The clinic had been closed since the spring flood took the east wing. It reopened with half its staff and no records. The roof leaked whenever the wind came off the river.

Only make that kind of change where you are content for the audio to differ from the print. If the audiobook is sold as unabridged, heavy rewriting is a decision about the edition, so keep it to what the ear needs.

Consistency

Pick one spelling for each character, place and term and apply it throughout. Check dialect spellings and phonetic dialogue, which a voice may read literally. Check that curly and straight quotation marks are used consistently, since dialogue boundaries depend on them. If you plan several voices, clear speaker tags matter more; see multi-voice fiction.

Build the pronunciation sheet now, while you are searching the text. A simple table is enough, and these entries are invented:

The last column exists because ElevenLabs says its dictionary searches are case sensitive. Its docs do not say how possessives are matched, so check one by listening.

When you are done

Export a clean copy, keep the original, and do a short trial on one chapter before committing the whole book. Pick the chapter with the most names, numbers and dialogue, not the first one. On that chapter, confirm that:

• the chapter title is read once, in the right place;
• no symbol, footnote number or web address is spoken by accident;
• the scene break produces a gap you can hear;
• each name on your sheet comes out the way you wrote it.

If you have not chosen a voice yet, run a same-chapter comparison on this cleaned text. Then follow the full workflow, and use the proofing checklist once the audio exists. The AI audiobooks hub lists related guides.

Last verified 2026-11-08. Rules and prices change: check the linked primary page before you rely on this.