Discussion Paper: Language Switching in Digital Publications

Status: Public Discussion Draft

Date: July 6, 2026

Status of This Document:

This document is intended as a discussion starter for the publishing, accessibility, and reading system communities. It does not propose recommendations or best practices. Instead, it identifies questions, challenges, and areas for further research related to language switching in EPUB publications and digital reading environments. The document is based on discussions within the DAISY Consortium Transition to EPUB 3 group and is intended to encourage broader community feedback before future best practices are considered.

1. Introduction

Modern publications sometimes contain multiple languages within a single work. Examples include quotations, technical terminology, proper names, educational materials, scholarly works, multilingual publications, and historical texts containing languages such as Latin or Ancient Greek. In EPUB and HTML, language markup can identify changes in language within content. Such markup may affect screen reader pronunciation, Braille translation, Read Aloud behavior, typography and hyphenation, generated audio, and the overall reading experience. At the same time, participants identified that proper implementation requires resolving several questions around when language switching markup is appropriate, how granular such markup should be, how users benefit from it, and how practical it is within publishing workflows.

2. Accessibility Requirements and Open Questions

The discussions frequently referenced WCAG’s “Language of Parts” success criterion. However, participants identified uncertainty around what constitutes a “part” of language. Questions raised included:

Does a “part” mean a single word, phrase, sentence, or larger section?
Should borrowed words be marked?
Should proper names be excluded?
How should titles and technical terms be treated?
Should language-learning publications be handled differently?

Participants noted that the granularity of markup has major implications for publishers and producers of accessible publications because fine-grained tagging can require substantial work.

Several participants also noted that multilingual publishing realities may differ significantly across languages and regions, especially outside primarily English-speaking publishing environments.

No consensus was reached on how “Language of Parts” should be interpreted in practice.

3. Typographic and Editorial Conventions

Participants generally agreed that existing print publishing conventions should be studied before developing EPUB-specific guidance. The discussions referenced editorial resources such as:

The Chicago Manual of Style
The Cambridge Handbook on Copy Editing Participants noted that these resources distinguish among language-learning materials, proper names, technical terminology, quotations, titles, and borrowed words.

Examples discussed included italicization of foreign words, glossary entries, dictionary-style definitions, and imported words that have become part of another language.

Some participants suggested that typographic distinction may help indicate where language markup is appropriate. Others questioned whether visual styling alone should influence accessibility markup.

No consensus was reached.

4. Borrowed Words and Common Usage

One recurring issue involved imported or borrowed words, sometimes referred to as “loan words,” that have become integrated into another language. Examples discussed included “zeitgeist”, “weekend”, and French usage of words such as “parking”.

Several participants generally agreed that commonly adopted words which are normally pronounced using the primary language of the publication may not require language switching markup.

A broader discussion theme that received support was the idea that human narration practices may provide a useful conceptual model for determining when language markup is appropriate.

Participants noted that:

if a human narrator would naturally switch pronunciation, accent, or language while reading content aloud, language markup may be appropriate;
if a human narrator would read the word naturally within the primary publication language, markup may not be necessary. Participants connected this idea to text-to-speech and Read Aloud systems because properly marked-up content may allow reading systems to switch language behavior at the appropriate point.

At the same time, participants noted that:

different readers may pronounce the same borrowed word differently,
multilingual switching may become disruptive,
and context may matter significantly.

5. Production Challenges for Publishers

Production burden was one of the strongest recurring themes in the discussions. Participants noted that:

publishers often do not currently perform granular language tagging,
authoring workflows are not optimized for it,
and fine-grained tagging of words and short phrases may require many hours of work.
The production is often outsourced by the publishers, and it is observed that the employees of the vendor may not know the language appearing in the publication, consequently it become challenging for them to identify language changes.
Participants also noted that bibliographies, references, and footnotes may present particular production challenges. A single bibliographic entry may contain titles of works, journal names, publisher names, or other information that should remain in the language of the cited resource, while editorial text added by the publication—such as "available online", "accessed", or similar phrases—remains in the primary language of the publication. Other components, such as publication years, page numbers, identifiers, and URLs, are language-neutral. Determining which portions of an entry should be identified with language markup may therefore require particularly fine-grained tagging.
Current authoring tools may also lack support for some language codes. Examples discussed included Latin and Ancient Greek in Microsoft Word and Adobe InDesign, as well as languages such as Wolof, where support across authoring tools and assistive technologies may be limited. Participants generally agreed that such deficiencies should be reported to tool developers.

6. Dead Languages and Unsupported Languages

The discussions explored cases involving Latin, Ancient Greek, and languages with limited text-to-speech or Braille support. Participants questioned:

whether unsupported languages still provide user benefit when correctly marked,
whether substitute language codes are appropriate, for example selecting the closest available supported language code when the correct language is not supported,
and how future AI-based test-to-speech may change expectations. At the level of principles, several participants generally agreed that correct language markup may still have long-term value even when current technologies lack full support.

One example raised during feedback discussions was the use of a Spanish voice for Māori text when a dedicated Māori voice is unavailable because it may provide the closest pronunciation.

At the same time, participants emphasized that production realities and tool limitations must also be considered. Several possible approaches were discussed:

Leave content in the primary document language.
Use a supported language code that approximates the intended language.
Use the technically correct language code even if support is incomplete.

No agreement was reached regarding which approach is preferable on a practical basis.

7. Screen Readers, Braille, and Read Aloud

Participants discussed how language markup affects screen reader pronunciation, automatic voice switching, language announcements, Braille translation, and Read Aloud behavior.

Some participants noted that screen readers may:

announce language changes,
switch voices automatically,
or potentially support navigation by language changes in the future. Participants also noted that Braille users may benefit from language markup because translation tables and contractions may depend on language identification.

At the same time, several participants observed that frequent voice or dialect switching can become disruptive for some readers.

Participants also observed that language changes involving a different writing system may present additional accessibility challenges. Unlike languages that share the same script, content written in scripts such as Arabic or Japanese inside an English publication may not be pronounced correctly—or may not be spoken at all—if appropriate language mark-up is unavailable.

Participants generally agreed that user preferences and personalization are important considerations.

8. Human Narration Practices

Participants discussed whether professional narration practices used in DAISY audio books and audiobooks may provide useful guidance for language switching behavior in digital publications.

Participants noted that human narrators naturally distinguish between:

foreign-language phrases intended to be pronounced differently,
and imported words that are spoken naturally within the primary language of the publication.

Several participants generally agreed that this distinction may help determine when language markup is appropriate.

The discussions also raised questions about whether generated audio and Read Aloud systems should attempt to emulate human narration behavior where possible.

No consensus was reached regarding how closely automated systems should follow human narration practices.

9. Questions for Community Feedback

The discussions revealed that language switching in EPUB/HTML publications touches many different areas of publishing and accessibility, often with competing needs and expectations.

Participants discussed language switching primarily in the context of language information provided through publication markup. Different technologies may use this information in different ways. For example, screen readers may use it to influence pronunciation and voice switching, refreshable Braille displays may use it to apply language-specific translation rules, and Read Aloud systems may use it to influence generated speech. The discussions identified a number of open questions regarding the expected behavior of these technologies and the markup practices needed to support them.

Expected Behaviour of Reading Technologies: