Pronunciation Lexicons Bind a Written Form to a Spoken Form

Suppose an assistant reads a purchase confirmation for a fictional product named Veyra. The interface shows the right spelling, but the voice stresses the wrong syllable. Adding another instruction to the assistant’s prompt may change the generated text; the speech engine still needs a pronunciation for the written name.

The W3C Pronunciation Lexicon Specification (PLS) defines an XML document that maps written forms to pronunciations for speech synthesis and recognition. Each lexeme groups one or more grapheme elements with one or more phoneme or alias elements. Here, grapheme contains the written word or short phrase. It does not mean a single Unicode grapheme cluster. A minimal entry for the intended English pronunciation “VAY-ruh” could be:

<?xml version="1.0" encoding="UTF-8"?>
<lexicon version="1.0"
  xmlns="http://www.w3.org/2005/01/pronunciation-lexicon"
  alphabet="ipa" xml:lang="en-US">
  <lexeme>
    <grapheme>Veyra</grapheme>
    <phoneme>ˈveɪrə</phoneme>
  </lexeme>
</lexicon>

An SSML document can reference the PLS file through its <lexicon uri="..."/> element. Creating the XML file alone does not make a speech engine load it.

The root’s alphabet chooses the default phonetic notation, and xml:lang identifies one language for that lexicon. PLS requires conforming processors to support ipa, the International Phonetic Alphabet. A phoneme can override the default with a vendor-defined x-... alphabet, but that choice ties the entry to processors that understand it. If an application speaks several languages, separate lexicons can supply language-specific pronunciations.

An alias takes a different route: it substitutes another written form, useful for an acronym whose expansion should be spoken. The processor then determines how to pronounce that substitute. Use phoneme when the sound itself needs explicit control. Multiple spellings can share a lexeme’s pronunciations, and multiple pronunciations can be listed. For speech synthesis, prefer="true" selects a preferred one; without it, document order applies unless the processor documents its own selection method. Recognition processors consider the listed alternatives valid.

The specification defines processor behavior. It does not guarantee that a particular voice service loads PLS or produces identical audio. Check the service’s lexicon interface, use the locale actually sent to the voice engine, and listen to the resulting name in representative sentences. Keep the spelling and pronunciation together in a maintained lexicon so a product rename or new locale has one clear place to review.