EPUB to TXT Converter

The conversion is performed by Calibre's command-line tool ebook-convert, invoked on the server as ebook-convert <input.epub> <output.txt> with --txt-output-encoding utf-8 so the result is always UTF-8 regardless of the original declaration. The pipeline has three stages.

First, the EPUB is parsed. An EPUB is a ZIP archive whose first entry is an uncompressed mimetype file containing application/epub+zip, followed by META-INF/container.xml pointing at the root .opf package file. The .opf declares metadata (Dublin Core elements), a manifest listing every resource, and a spine defining reading order. The parser resolves the spine, then loads each XHTML content document in turn. This is why the text comes out in reading order rather than archive order — the spine is what defines the sequence, not the filenames.

Second, the XHTML is reduced to text. The DOM of each content document is walked and only text nodes are kept, so every element — <p>, <h1>, <em>, <span>, <div> — contributes its character data and nothing else. Block-level elements are converted into paragraph breaks, which is what produces the blank lines between paragraphs in the output; inline styling is simply dropped. Structural elements with no text content, such as <img> and <hr>, vanish entirely. Links lose their target but their anchor text remains in place, so URLs written as visible text survive while a linked word loses the address behind it.

Third, the result is serialised. Whitespace is normalised, the text is written out in the target encoding, and the file is finished. Two consequences are worth flagging. Any content that exists only as a picture is gone: a scanned novel, a comic in fixed-layout EPUB (rendition:layout=pre-paginated), or a children's picture book has essentially no text nodes, so the output is empty or nearly so. And no markup means no structure — CSS page-break rules, which the PDF pipeline uses to paginate, are meaningless here, so chapter boundaries survive only as whatever heading text the heading itself contained.

• Engine: Calibre ebook-convert, run server-side as ebook-convert <in.epub> <out.txt> --txt-output-encoding utf-8
• EPUB container: ZIP with an uncompressed mimetype (application/epub+zip), META-INF/container.xml pointing to the root .opf, and the .opf carrying metadata / manifest / spine
• Reading order comes from the spine, not from filenames or ZIP order — this is why extracted chapters appear in the intended sequence
• Text extraction keeps text nodes only: block elements become paragraph breaks, inline styling is discarded, and elements without character data (<img>, <hr>) disappear
• Lossy by design: images, cover, tables (cells serialised as running text), fonts, colours, and TOC hierarchy do not survive; links keep only their anchor text
• Image-only sources fail: a scanned EPUB or a fixed-layout comic has no text nodes, so the output is empty — the service detects the empty result and reports a failure instead of handing you a blank file
• Output is UTF-8 plain text with normalised whitespace; no page concept, so page numbers and pagination settings have no effect