Text to audiobook: EPUB or PDF to MP3.

At a glance

How an EPUB or text file becomes an audiobook

An EPUB is a zip of web pages plus a list, the spine, that gives their reading order. The page reads that list, turns each document into plain paragraphs, drops the cover, copyright and contents pages, and names each chapter from its first heading. A text file is split wherever a line looks like a heading: "Chapter 12", "CHAPTER XII", "Part Two", "Prologue", or a Markdown # or ##. A long text with no headings at all is cut into parts of about six thousand words at paragraph breaks. You can untick chapters (an index, the acknowledgements) and rename any of them before you start.

Each chapter is then read sentence by sentence by Kokoro, an open neural text-to-speech model, on your GPU if the browser can use it and on the CPU otherwise. Numbers, prices, years and abbreviations are written out first, so "1847" is read as a year and "Dr." as "Doctor". As soon as a chapter is finished it is compressed — AAC for the M4B, MP3 for the zip — and saved in the browser. A long book never has to sit in memory as raw audio, and an interrupted run carries on from the last finished chapter instead of from page one.

At the end, the chapters are joined without re-encoding and the M4B gets a chapter track with your titles, plus the book title and author as tags. Apple Books, VLC, Smart AudioBook Player on Android, BookPlayer on iOS and most dedicated audiobook apps show those chapters and remember your place.

PDF to audiobook

A PDF is a picture of a page, not a book: it stores where each run of letters is drawn, with no paragraphs, no reading order and no idea what a heading is. So the page rebuilds them. Runs at the same height become lines, a gap of more than a line ends a paragraph, a word hyphenated at the end of a line or a page is joined back up, and the running head and page number that repeat at the top or bottom of every page are dropped. A two-column layout, as in most academic papers and many magazines, is read down the left column and then the right, not straight across the gutter. Figure and table captions are skipped, because read aloud they land in the middle of a sentence, and so are figure labels and footnotes in smaller type. Chapters come from the PDF's own bookmarks when it has them; otherwise from headings in larger type, numbered sections such as "3. Results", and lines like "Chapter 7". On a real two-column paper (He et al., ResNet, CVPR 2016) it found Abstract, the four numbered sections and References, so you can untick the references before you start.

What a PDF cannot give is text it does not contain. A scanned book is a stack of photographs of pages, and the page says so rather than reading nothing; run it through OCR first (Adobe Acrobat, Google Drive's "Open with Google Docs", or the text recognition in macOS Preview) and try the result. Password-protected PDFs need the password removed first. Equations, tables and code are read as the characters they contain, which rarely sounds like anything, so for a textbook expect to skip those parts. An EPUB of the same book, where one exists, always reads better: it carries the real paragraphs and chapters.

How long it takes to make an audiobook with AI

Speech runs at roughly 150 words a minute, so a 90,000-word novel is about ten hours of audio. On a laptop with a GPU that the browser can use (current Chrome or Edge on a desktop), making it takes about half as long as listening to it: a ten-hour novel in about five. On the CPU it takes several times as long. Either way a short story or an article is done in minutes, and a full novel is an evening's or an overnight job. The estimate under the chapter list is for your book and your machine. Keep the tab open; a laptop should be plugged in, because this works its processor hard. Because finished chapters are saved, a full book can be made over several sessions.

Choosing an AI narrator for a whole book

For long listening, the voices graded A and B in the list hold up best: Heart and Bella (American) and Emma (British) have the most natural rhythm over hours. Try "hear the opening" with two or three before committing to a whole book. 1.0× is the model's natural pace; most audiobook listeners speed up in their player, which keeps the pitch steady anyway, so there is little reason to bake a faster speed into the file. Every chapter is labelled in its tags as synthetic speech.

What it can't do

It reads, it doesn't act: one voice for the whole book, in a calm narrator's style, with no character voices and no shouting or whispering. Pronunciation of invented names is espeak-ng's best guess, which is usually reasonable and occasionally odd; respell a name in the text to fix it. Footnotes are skipped, and tables, captions and images are not described. EPUBs bought from Kindle, Apple Books or Kobo are usually DRM-protected and cannot be read by anything outside the store's app; the page says so if you drop one. Project Gutenberg, Standard Ebooks and most independent authors' EPUBs are DRM-free. Spanish, French, Italian, Portuguese and Hindi books can be read with a voice in that language, but numbers and abbreviations are only written out properly in English.

Respect copyright: making an audiobook of a book you own for your own listening is one thing; distributing it is another. Public-domain books are free for anything.