Skip to content

Fixing names and terms a synthetic voice gets wrong

Last checked November 2, 2026

A voice gets a name wrong because it has never met that name. It looks the word up, or sounds it out from the spelling, and a brand name made of invented letters falls into the gap.1

The fix is to tell the voice how the word should sound. There are several ways, and the simplest one usually works.

Why it happens

One approach to turning spelling into sound is a dictionary of words and their pronunciations. The source says that approach completely fails when it's given a word that isn't in the dictionary. Systems usually combine it with rules that sound words out, which is why an odd name often comes out as a confident, wrong guess.1

Amazon puts the same thing plainly. A text-to-speech engine reads the text literally, pronouncing the name exactly as it is spelled. Its example is a stylized name with numbers in place of letters.2

The fixes, simplest first

Respell the word in your script. Write the name the way it sounds, for example "Shiv-awn" for the Irish name Siobhan. It costs nothing and works in any tool. The catch is that the written script now looks odd, so keep a clean copy for the people who read it.

Use a substitute tag. SSML has a tag called sub. The text in its alias value replaces the contained text for pronunciation.3 Your normal spelling stays in the script, and the voice says the alias. The same tag expands acronyms, such as turning a short form into the full organization name.3 See what SSML is for the tags around it.

Spell it out letter by letter. For an acronym that should be said as letters, the say-as tag lets you tell the voice the type of text it's looking at.3

Use a pronunciation lexicon. A lexicon is a saved list of words and how to say them. Amazon says pronunciation lexicons let you customize the pronunciation of words. You can give an alias, or write the pronunciation in a phonetic alphabet.2

Fixing a wrong pronunciation, simplest firstListen and note the wrong wordRespell it in plain lettersUse a substitute tag if availableUse a pronunciation lexiconListen to the whole file again
Start at the top and stop as soon as the voice gets it right.

Things to check with a lexicon

Lexicons are the right tool when the same names come up again and again, such as in a long series of training audio. Read the maker's rules before you build one.

  • Amazon says lexicons must follow the W3C Pronunciation Lexicon Specification.2
  • A request can name a maximum of 5 lexicons.4
  • A lexicon is applied only if its language matches the language of the voice.4

Other makers have their own versions of this feature, with different names and limits. Look for "custom pronunciation" or "dictionary" in their documentation, and note the date you looked.

After any fix

Play the line again. Respelling sometimes shifts the stress to a different part of the word, and a fix for one name can leave the next sentence sounding odd. Then play the whole file, since small changes can alter the pauses nearby.

Keep a short list of your tricky words and the fix that worked, so the next script goes faster.

Where the facts come from

  1. Speech synthesis, Wikipedia
  2. Managing lexicons, Amazon Polly documentation
  3. Speech Synthesis Markup Language (SSML), Google Cloud Text-to-Speech documentation
  4. SynthesizeSpeech, Amazon Polly documentation