← Blog · 16 July 2026

On ordinary text, good synthetic voices are now indistinguishable from a human one. The last obstacle is the proper noun: a rare surname, a local place name, a regional acronym. A voice that reads an editorial flawlessly can mangle the mayor's name on the next line — and for a regional title, that's precisely the name that matters.

Why proper nouns resist

A speech engine predicts pronunciation from regularities learned on huge corpora. Those regularities work for common words, but a local proper noun doesn't obey the general rules: many place names aren't read as they're spelled, some acronyms have to be spelled out, others expanded. The model makes a reasonable guess — and gets it wrong exactly where it's most noticeable.

The wrong fix: regenerate everything

You might hope the model would "learn" your names. But retraining an engine for a handful of towns makes no sense, and hand-correcting each audio would ruin the whole point of automation. You need a correction that lives alongside the text, not inside it — durable, reusable, with no new generation.

The right fix: a pronunciation lexicon

With WeDispatch, you keep a small lexicon, one line per correction, in the form name = pronunciation. It's applied before each synthesis, on all following audios — including in charter mode, because it's a rule, not an AI. A sub-editor can fill it in between deadlines, with no technical skill.

A detail that matters: the displayed text — highlighting, subtitles — follows the corrected pronunciation, so the word shown on screen is exactly the one you hear.

A matter of credibility, not comfort

For a regional daily, pronouncing local names correctly isn't cosmetic: it's a question of credibility with readers who know these places better than anyone. A pronunciation lexicon turns the historic weak point of text-to-speech into a non-issue. See the solution in action on the regional press page.