The best file format for AI narration, ranked

Reflowable ebook formats carry chapters, headings and metadata. Reading order is explicit, so chapter lists are accurate and front matter is easy to separate. If you have the choice, use this.

Fix before uploading: nothing, usually. If the file is one giant chapter, add headings in Calibre’s editor.

Word’s heading styles map cleanly to chapters and body text is unambiguous. Headers, footers and comments are not narrated.

Fix before uploading: use Heading 1 for chapter titles; accept or reject tracked changes; keep text out of floating boxes.

Plain text narrates flawlessly at the sentence level; the only thing missing is structure. Chapter headings are detected from lines such as “Chapter 3” or “# Part Two”; otherwise the text is split into even sections. HTML keeps its headings, but menus and footers are text too.

Fix before uploading: put chapter titles on their own line; add “Title:” and “Author:” lines at the top; save as UTF-8.

Usable, with work: page numbers and running headers are stripped, hyphenation undone, paragraphs re-joined across pages, chapters taken from the outline or heading sizes. Two-column academic layouts and heavy tables are where it strains.

Fix before uploading: if you also have the EPUB or DOCX, use that. Otherwise export a PDF with an outline (bookmarks) from the source application.

A scan without a text layer has nothing to narrate and is rejected. Scans with OCR text narrate like any PDF, with OCR errors read aloud. Comics (CBZ, CBR) and DJVU go through OCR and produce fragmentary text.

Fix before uploading: run OCR in your scanner app, or find a text edition.

Files are limited to 5 MB. Text is small: a 100,000-word novel is under 1 MB as EPUB. Anything larger is images; strip them and the same book fits.