EPUB to Text: Turning a Shelf of Books Into Something You Can Compute On
Books are the best-edited long-form prose we have. Professionally written, professionally copyedited, thousands of words of consistent voice with no navigation menus, cookie banners, or comment sections mixed in. As raw material for anything language-shaped — embeddings, style analysis, retrieval, training a classifier to recognise tone — that's hard to beat.
The problem is that it's sitting inside a zipped bundle of XHTML files. An EPUB to text converter gets it out: one continuous prose stream, no tags, no markup characters, nothing your tokeniser has to be taught to ignore. Upload above and it comes back in seconds. Free, no sign-up, 50 MB limit, nothing stored.
Why Flat Text Is the Right Shape for This Work
Plain text isn't the budget option here. For most things you'd do with a book at scale, it's the format the tooling actually expects:
- Embeddings and semantic retrieval. Chunk the prose, embed each chunk, retrieve by similarity. Markdown hashes and asterisks occupy token positions and contribute nothing to meaning — they dilute the vector. Clean paragraphs retrieve better, and every chunking library on earth takes text by default.
- Readability and stylometry. Flesch-Kincaid, average sentence length, type-token ratio, vocabulary richness, punctuation rhythm. These are word- and sentence-level measures, and formatting tokens corrupt every one of them.
- Search indexing across a library. Postgres
tsvector, SQLite FTS5, Elasticsearch, or a plain folder yougrep. All of them want raw text in a field. Fifty converted books make a genuinely fast personal search engine. - Corpus linguistics and NLP pipelines. Concordancing, collocation analysis, named entity extraction, topic modelling. spaCy and NLTK take strings, not markup.
- Text-to-speech and accessibility. A TTS engine handed Markdown will cheerfully read "hash hash Chapter Four" out loud.
- Diffing editions. Two printings of the same title, extracted to text, run through a word-level diff — that's how you find what a second edition actually changed.
What You Give Up by Flattening
Chapter boundaries. That's the real cost, and it's a bigger deal for books than for most document types, because a book is long enough that "where am I" matters.
In the source EPUB, each chapter is a separate XHTML file with a heading tag and an entry in the navigation document. Flattened, a chapter title becomes a short line of text sitting above a long line of text, with nothing marking it as different. Ask a model to "summarise chapter seven" and it has to work out where seven starts from the wording alone. Usually it manages. Sometimes it starts a paragraph late and quietly summarises the wrong thing.
Blockquotes stop being visually distinct from narration too, which matters for non-fiction where quoted source material and the author's own argument sit side by side. If any of that is load-bearing for what you're doing, take EPUB to Markdown instead — the heading tree survives there and you can slice the file cleanly by chapter. The format toggle at the top of this page switches between them and carries your selected file across, so you can produce both and compare without uploading twice.
Typography Artefacts Worth Cleaning Up
Publishers typeset properly, which is great for reading and mildly annoying for computation. A handful of things reliably show up in extracted book text, and a single pass of find-and-replace handles most of them:
- Curly quotes and apostrophes. Books use typographic quotes, not
"and'. Tokenisers usually cope, but naive regex and string matching won't — search fordon'tin a novel and you may get nothing while the page is full of don’t. - Em dashes and en dashes. Often unspaced, which means sentence splitters occasionally glue two sentences together or break one in the wrong place.
- Soft hyphens and non-breaking spaces. Invisible on screen, present in the byte stream, and quietly responsible for word counts that don't match anything.
- Ligatures. "fi" and "fl" are sometimes single glyphs. They normally extract fine, but an unusual embedded font can produce odd characters in the odd word.
- Front and back matter. Title page, copyright notice, dedication, acknowledgements, "about the author", index remnants. Boilerplate that shows up in every book you convert, so it becomes spuriously common vocabulary across a corpus. Trim it if you're doing frequency work.
None of this is the converter being sloppy. It's what's genuinely in the file, and knowing to look for it is most of the battle.
DRM, and Where Convertible Books Come From
A book bought from a major retailer is usually encrypted. The XHTML inside the container is scrambled and
an encryption.xml file declares it, so no converter — this one included — can read a word of
it. Upload one and you'll get an error or an empty output. Worth knowing before you spend ten minutes
wondering what went wrong.
Plenty of good sources are DRM-free by design. Project Gutenberg's public domain catalogue is enormous and ideal for corpus work. Standard Ebooks produces carefully typeset public domain editions with unusually clean markup. Most technical publishers, arXiv-adjacent presses, Creative Commons titles, and anything you exported yourself all convert without complaint. Converting books you own for your own study is ordinary use; redistributing the extracted text of a copyrighted book isn't. Enough said.
Doing This Across a Whole Library
Converting one book is a two-second job. The interesting version is converting forty and treating the result as a dataset. A few things that make that go smoothly:
- Zip the batch. ZIP to text takes an archive and converts every supported file inside it in one pass, which beats forty separate uploads. Watch the 50 MB ceiling — EPUBs are small, but illustrated ones aren't.
- Normalise before you chunk. Straighten quotes, strip soft hyphens, collapse repeated blank lines. Do it once at ingest and every downstream step gets simpler.
- Chunk on paragraph boundaries, not fixed character counts. Books have real paragraphs and they're natural semantic units. Blind 1,000-character splits cut sentences in half and produce embeddings for fragments.
- Keep the filename as metadata. Once it's all one text stream, the title and author are the only provenance you have. Store them alongside each chunk or you'll retrieve a great passage and have no idea which book it came from.
- Check the token count. The counter under the output gives you the real number for a given title, which is the difference between planning a pipeline and guessing at one.
EPUB Is Better Source Material Than the PDF Version
If the same title exists in both formats, the EPUB gives you cleaner text with less work. A PDF is fixed-layout: every visual line is a separate run of glyphs, so paragraphs arrive broken into short lines, words hyphenated across line breaks stay hyphenated, and the running header and page number land in your prose every few hundred words. PDF to text deals with all of it, but you'll be scrubbing afterwards.
EPUB is reflowable. There are no pages, therefore no page furniture, and paragraphs are actual paragraphs in the markup — so they come out as continuous lines that a regex search can match across. For anything involving a phrase search, sentence splitting, or embedding, that difference alone is worth the swap.
Frequently Asked Questions
How do I convert an EPUB to text?
Drop the .epub into the converter above and the book's text comes back as plain text you can download as .txt. Free, no sign-up, 50 MB per file — comfortably more than any reflowable e-book needs, since EPUB is compressed HTML and even long books rarely pass a few megabytes.
Does it work on books bought from Kindle or Apple Books?
Only if they are DRM-free. A purchased book is usually encrypted and tied to the account that bought it, and an encrypted file has no readable text inside for any converter to reach. Public domain titles from Project Gutenberg, Standard Ebooks, technical books from publishers who sell unencrypted files, and your own exports all convert normally.
Do chapters stay separated in the output?
Chapter boundaries survive as breaks in the flow, but chapter titles arrive as ordinary lines with nothing marking them as headings — that is the nature of flattening to text. If you need to navigate or chunk by chapter afterwards, EPUB to Markdown keeps the heading levels that make that possible.
What about footnotes and endnotes?
They come through, generally collected at the end of the section they belonged to rather than inline where the marker sat. For scholarly texts this means a run of note bodies appears between chapters. Easy to read past, worth knowing about if you are chunking the output for embeddings and wondering why one chunk is all citations.
Can I feed a whole book to an LLM this way?
You can convert it in one pass, but check the token count before pasting. A 300-page novel runs to roughly 120,000 tokens — inside a 200K context window, and comfortably over what a chat interface will accept in a single message. The counter under the output tells you the number before you find out the hard way.
The Rest of the Toolkit
EPUB is one of thirteen supported formats. File2Txt handles all of them from a single upload. For manuscripts there's Word to text, for legacy rich-text documents RTF to text, and for saved articles HTML to text. Structured data of your own goes through CSV to text or JSON to text.
Beyond documents: convert a GitHub repository to text, a GitLab project, or a local folder, and use Web2Txt to pull live pages down as text. There's a longer guide to preparing documents for LLMs if you want the wider picture.
Repo2Txt is built and maintained by v12hero, an independent developer building privacy-first native and web apps.