Roughly fifteen percent of the world reads with some form of disability. For that audience, an ebook is not an aesthetic preference but the only practical way to consume long-form text. The book that arrives with broken heading hierarchy, image-only diagrams, or a fixed-layout PDF that ignores font scaling is, for a print-disabled reader, a closed book. Accessibility in ebook conversion is therefore a craft, not a checkbox.

The good news is that the standards are mature, the tooling is open source, and a remediation pass that takes one day on a typical book yields a file that works for screen readers, refreshable braille displays, voice control, and large-print users alike. The bad news is that most pipelines still output files that fail at the first chapter heading. This article walks through the conversion workflow that actually produces accessible EPUB 3 output, the validators that catch real problems, and the editorial decisions that automated tools cannot make for you.

Why EPUB 3 Is the Target Format

EPUB 3 is the only mainstream ebook format whose specification was written with accessibility as a first-class concern. The format is essentially a zipped bundle of XHTML, CSS, and SMIL, governed by the Open Web Platform. Because the content layer is HTML, every accessibility lesson learned on the web since the WCAG 2.0 publication in 2008 transfers directly into the book.

The EPUB Accessibility 1.1 specification published by the W3C aligns conformance levels with WCAG 2.2 and adds publication-specific requirements: a declared accessMode, a accessibilitySummary, navigation that exposes both a logical table of contents and a page list when print pagination is preserved, and metadata describing whether the work has been remediated for math, image descriptions, or sign language video.

FormatReflowable TextSemantic MarkupScreen Reader SupportPrint Disability Standard
EPUB 3YesXHTML5 with ARIA rolesStrong on iOS, Android, JAWS, NVDAW3C EPUB Accessibility 1.1
PDF (untagged)NoNone reliableLinear text only, often brokenAvoid for new content
PDF (PDF/UA)NoTagged treeAdequate when authored correctlyISO 14289
MOBI / KF8YesSubset of HTMLLimited, Amazon ecosystemDeprecated in 2022
DAISY 3YesSpecialized DTBookExcellent for synchronized audioANSI/NISO Z39.86
Plain HTMLYesXHTML5StrongWCAG 2.2 directly
The MOBI format is no longer accepted by Amazon for new uploads as of 2022, and KF8 ingests EPUB internally. Plain HTML is excellent technically but fails the offline reading and rights management expectations of book buyers. PDF remains common because layout designers love it, but a PDF that has not been carefully tagged is effectively unreadable by any assistive technology.
"Accessibility is a quality of the user's experience, not a feature you add at the end. If you cannot test it the day you draft the manuscript, you cannot guarantee it the day you publish." Sarah Horton, A Web for Everyone

The Conversion Pipeline That Actually Works

A conversion pipeline that produces accessible EPUB does four things in order: it preserves semantic structure from source, it adds missing alternative descriptions, it validates the package, and it tests with at least one real screen reader before release. Skip any step and the output regresses to the lowest common denominator.

The source matters more than the converter. A Word manuscript that uses paragraph styles named Heading 1, Heading 2, Caption, and Block Quotation can be transformed into clean EPUB by Pandoc in a single command. A Word manuscript that uses bold sixteen-point Calibri to indicate chapter titles produces an unstructured stream that no converter can untangle.

pandoc manuscript.docx \
  --from=docx+styles \
  --to=epub3 \
  --metadata title="The Quiet Garden" \
  --metadata author="A. Reader" \
  --metadata lang=en \
  --epub-cover-image=cover.jpg \
  --toc --toc-depth=2 \
  --css=accessible.css \
  --output=quiet-garden.epub

The --from=docx+styles flag tells Pandoc to honor the paragraph styles rather than guess at structure from formatting. The lang metadata is non-negotiable because screen readers switch pronunciation engines based on it; a German section in an otherwise English book without an explicit xml:lang attribute will be read aloud as if it were English, which is incomprehensible.

For books with complex apparatus, footnotes, indexed terms, and cross-references, the dedicated tooling around Calibre or the commercial Vellum on macOS produces cleaner output than Pandoc. Both can be driven from the command line for batch jobs.

Semantic Structure Is Eighty Percent of the Work

Screen reader users navigate by structure, not by visual position. They press a single key to jump to the next heading, the next list, the next table, the next link. A book without semantic structure forces them to listen linearly from page one, which on a four-hundred-page novel means hours of unskippable preamble.

The minimum semantic structure for any prose ebook covers six elements: chapter and section headings in proper hierarchy, paragraphs marked as paragraphs, lists marked as lists, blockquotes marked as blockquotes, page breaks marked with the EPUB epub:type="pagebreak" attribute when print pagination is preserved, and language attributes wherever the language changes. EPUB 3 adds publication-specific roles via epub:type for chapters, prologues, footnotes, glossaries, and bibliographies.

<section epub:type="chapter" role="doc-chapter" aria-labelledby="ch3-title">
  <h1 id="ch3-title">Chapter Three: The Map Maker</h1>
  <p>The first map Octavia drew was not of land but of silence...</p>
  <aside epub:type="footnote" role="doc-footnote" id="fn-12">
    <p>Mercator died before completing the projection that bears his name.</p>
  </aside>
</section>

The role attributes from the DPUB-ARIA module make these structures legible to assistive technologies that do not yet understand epub:type. Both should be present until adoption of EPUB-only roles is universal, which it currently is not.

Image Descriptions: The Hardest Part

Alternative text is the single biggest editorial decision in accessibility remediation, and it cannot be automated. Machine-generated alt text from vision models is improving but is not yet trustworthy for technical diagrams, art reproductions, or data visualizations. The descriptions need to be written by someone who understands what the image is doing in the argument of the book.

The W3C distinguishes four cases. Decorative images get alt="" and role="presentation". Simple informative images get a short alt attribute. Complex images such as charts, infographics, or maps need a long description, either inline or via the aria-describedby mechanism. Images of text need the text transcribed, never paraphrased.

Image TypeAlt Text LengthLong Description NeededTechnique
Decorative ornamentEmptyNoalt="" and role="presentation"
Author photo5 to 15 wordsNoInline alt attribute
Simple chart15 to 30 wordsSometimesInline alt plus caption
Complex infographicBrief summaryYesaria-describedby to hidden div
MapBrief summaryYesLinked text alternative
Data table rendered as imageNone acceptableReplace with real tableHTML table markup
Mathematical equationNone acceptableUse MathMLEPUB 3 supports MathML inline
Image of text in a posterFull transcriptionNoTranscribe verbatim
Mathematical content deserves special attention. EPUB 3 supports MathML natively, and screen readers including JAWS, NVDA, and VoiceOver can read MathML aloud with appropriate pronunciation rules. A book that renders equations as PNG images forces blind mathematicians to either guess from context or abandon the work. Conversion tools like MathJax, Pandoc with the `--mathml` flag, and the Texas Instruments tools for converting LaTeX produce usable MathML output.
"An image without alternative text is a closed door. An image with bad alternative text is a door with the wrong sign on it. The reader prefers the first because at least they know to keep walking." Lรฉonie Watson, Accessibility for Everyone

Validation: Run These Two Tools Every Time

Two open-source validators handle most real defects. EPUBCheck enforces the technical specification: malformed XML, broken internal links, missing manifest entries, illegal nesting. Ace by DAISY enforces accessibility specifically: missing alt attributes, undeclared language, low-contrast inline color, missing page list when pagination is declared, missing accessibility metadata.

java -jar epubcheck.jar quiet-garden.epub
ace --outdir reports quiet-garden.epub

EPUBCheck failures are blockers for distribution. Ace findings are graded by severity; a clean Ace report with zero serious or critical findings is the minimum bar for publishers serving educational or government markets in the European Union, where the European Accessibility Act became enforceable on 28 June 2025.

Validators do not catch every problem. They cannot verify that the alt text describes the image accurately, that the heading hierarchy reflects the document's logical structure, or that the reading order matches authorial intent on a multi-column layout. Manual review remains essential.

Testing With Real Assistive Technology

The final and most often skipped step is opening the file in a real screen reader and listening to a chapter. Five minutes of NVDA on Windows or VoiceOver on iOS catches problems no validator will surface: a chapter heading announced as a paragraph because of a level skip, a footnote read inline in the middle of a sentence, a language switch missed by the engine and pronounced phonetically, an image whose alt text describes the visual rather than the meaning.

The free combinations to test cover the major user populations: NVDA with Firefox or Edge on Windows, VoiceOver with Apple Books on iOS, TalkBack with Google Play Books on Android, and Thorium Reader on desktop, which is the W3C reference implementation for EPUB 3 accessibility.

For a reading platform, accessibility is also a UX research subject. The cognitive considerations explored at What's Your IQ on how working memory shapes comprehension apply directly to ebook design decisions about chunk size and navigation depth. The document workflow guides at File Converter Free cover the upstream conversions, DOCX to EPUB, that determine how clean the input to remediation will be.

Common Failure Modes and Their Fixes

Most accessibility regressions in ebook production trace back to a small set of recurring mistakes. Naming them helps editorial teams fix them once rather than repeatedly.

The first failure is treating headings as a typographic effect. When chapter titles are styled directly with bold and font size rather than tagged as Heading 1, the export tool sees them as ordinary paragraphs. Fix this in the source manuscript with paragraph styles, not in the EPUB output.

The second failure is missing language declarations on the package level. EPUB requires a <dc:language> element in the OPF metadata. Without it, screen readers fall back to the system locale, which for international readers is wrong by default.

The third failure is inline color used to convey meaning. Red text marking footnote references is invisible to color-blind readers and meaningless to screen readers. Use shape, position, and explicit text instead.

The fourth failure is fixed-layout EPUB used for prose. Fixed-layout EPUB has legitimate uses in cookbooks, picture books, and graphic novels, but for novels and reference works it removes the user's ability to scale text. Use reflowable EPUB unless layout is genuinely load-bearing.

The fifth failure is inadequate metadata. EPUB Accessibility 1.1 requires several metadata declarations including schema:accessMode, schema:accessibilityFeature, schema:accessibilityHazard, and a human-readable schema:accessibilitySummary. These appear in the OPF and are exposed by accessibility-aware reading systems and library catalogs. The educational platforms exploring exam preparation accessibility at Pass4Sure demonstrate how this metadata drives reader-side feature discovery.

Conformance Levels and What They Mean for Procurement

EPUB Accessibility 1.1 defines two conformance levels, A and AA, mirroring WCAG. Level A is the minimum any educational or government procurement contract should accept. Level AA is the working standard for major trade publishers and the requirement under the European Accessibility Act for ebooks sold to consumers in the EU.

Conformance LevelWhat It GuaranteesTypical Use
EPUB Accessibility 1.1 Level AStructural navigation, alternatives for non-text content, language declared, no flashingMinimum for any new publication
EPUB Accessibility 1.1 Level AAAdds image descriptions for complex graphics, MathML for math, page list for pagination, metadata completeRequired for EU market, US Section 508, most education markets
Level AAAAdds extended audio descriptions, sign language, reading level metadata, pronunciation hintsSpecialized accessible editions
No conformance claimUntestedAvoid; cannot be procured by accessibility-aware buyers
A book that claims conformance must also include a machine-readable certifier statement in metadata: who tested the book, against what standard, and on what date. This is enforceable by accreditation bodies and forms the audit trail for procurement disputes.

A Practical Remediation Workflow

A typical remediation pass on a 90,000-word novel that arrived as a Word file with mixed styling takes one editor three to six hours. The workflow that consistently completes in that window proceeds in five passes.

The first pass cleans the source: every chapter title gets Heading 1, every section gets Heading 2, every block quote gets the Quote style, every list uses Word's list facility rather than typed asterisks, every footnote uses Word's footnote feature, and every figure caption uses the Caption style. This is editorial work, not technical work.

The second pass runs Pandoc or the publisher's chosen converter with explicit metadata flags. Output goes to a working directory where the EPUB is unzipped for inspection.

The third pass adds image descriptions. Each image is reviewed against its role in the text; the editor writes alt text and, for complex images, a longer description that lives in a hidden block referenced by aria-describedby. This is the slowest pass and the one most often shortcutted, with predictable results.

The fourth pass adds metadata: accessibility mode, hazard declaration, summary written for end users, certifier identification, and the conformance claim with date. The OPF file is edited directly or generated by the publisher's metadata tool.

The fifth pass runs EPUBCheck and Ace, fixes any findings, then opens the result in NVDA and VoiceOver to listen to the first chapter, a chapter from the middle, the table of contents, and any complex figure. Findings from this listening pass go back into the source manuscript so they do not recur in future titles.

Tooling That Earns Its Place

A short list of tools covers nearly all production needs. Sigil for hand-editing EPUB, Pandoc for converting from Word and Markdown, Calibre for batch conversions and reader testing, Ace by DAISY for accessibility validation, EPUBCheck for spec validation, MathJax or Pandoc-MathML for equations, and Thorium Reader for reference reading. Commercial tools including Adobe InDesign with the EPUB export plugin and Vellum on macOS produce competitive output but require the same editorial discipline upstream.

For organizations producing many titles, the Daisy Consortium maintains the EPUB Accessibility 1.1 Knowledge Base and a free training program. The DAISY Accessible Publishing Knowledge Base is the single most useful reference document for an editorial team starting an accessibility program.

The discipline of accessible publishing has parallels in other domains. The note-taking workflows discussed at When Notes Fly demonstrate the same principle in personal information: structure that machines can parse is structure humans can navigate later. A book is a long note that thousands of readers will navigate; the investment in structure pays back at scale.

What Changes When Accessibility Becomes Normal

Publishers who run a remediation pipeline for a year report a consistent surprise: the editorial discipline that produces accessible ebooks also produces ebooks that perform better in print, in audiobook production, and in translation. Clean structure is portable. A book whose chapters are real chapters, whose figures are real figures with real descriptions, and whose metadata is honest about the work transfers cleanly into every downstream pipeline.

The reverse is also true. A book whose heading hierarchy is broken cannot be reliably indexed, cannot be reliably converted to audiobook, cannot be reliably translated by professional translation tools, and cannot be reliably reformatted for print-on-demand. Accessibility remediation is, in practice, production hygiene.

"When you build for the user with the most constraints, you discover what the medium actually requires. Everyone else inherits the work." Kat Holmes, Mismatch: How Inclusion Shapes Design

For related guidance on conversion workflows, see the role of file conversion in enhancing accessibility and how to choose the best archive format for data storage.

References

  1. W3C. (2025). EPUB Accessibility 1.1. World Wide Web Consortium Recommendation. https://www.w3.org/TR/epub-a11y-11/
  1. W3C. (2023). Web Content Accessibility Guidelines (WCAG) 2.2. https://www.w3.org/TR/WCAG22/
  1. ISO. ISO 14289-1:2014 Document management applications, Electronic document file format enhancement for accessibility, Part 1: Use of ISO 32000-1 (PDF/UA-1). https://www.iso.org/standard/64599.html
  1. DAISY Consortium. (2024). Accessible Publishing Knowledge Base. https://kb.daisy.org/publishing/
  1. European Union. Directive (EU) 2019/882 on the accessibility requirements for products and services (European Accessibility Act). https://eur-lex.europa.eu/eli/dir/2019/882/oj
  1. Horton, S., and Quesenbery, W. (2014). A Web for Everyone: Designing Accessible User Experiences. Rosenfeld Media. https://rosenfeldmedia.com/books/a-web-for-everyone/
  1. Watson, L., and Kalbag, L. (2017). Accessibility for Everyone. A Book Apart. https://abookapart.com/products/accessibility-for-everyone
  1. ANSI/NISO Z39.86-2005 (R2012). Specifications for the Digital Talking Book. https://www.niso.org/publications/ansiniso-z3986-2005-r2012-specifications-digital-talking-book