Teach a voice one word, once, for every article you publish

There is a class of errors no synthetic voice will ever fix on its own, and switching tools does nothing about it: the words you have to know in order to say them right. The French town "Ploërmel" is said "Plo-air-mel". No rule guesses it, no model invents it, because nothing in the spelling points to it. You have to know the place. The client knows, the machine does not, and it is precisely for these cases that WeDispatch keeps a pronunciation lexicon per account.

The move, in one line

The pronunciation lexicon is a list you keep yourself, one entry per line, in the form spelling = pronunciation. On the left, the word as it appears in your articles. On the right, how to say it, written out phonetically in plain spelling:

• Ploërmel = Plo-air-mel
• Cysoing = See-zwan
• URSSAF = Urssaf

That is all. You add the word that snags, you write how it should sound, and the correction applies to the next synthesis. A sub-editor fills this in between deadlines with no technical skill at all: there is no international phonetic alphabet to learn, only a readable transcription you tune by ear.

Three properties that matter, and one that surprises

First, it is a rule, not an artificial intelligence. The replacement is deterministic: at every synthesis the same word gets the same pronunciation, across every later article, charter mode included. There is nothing to hope for and nothing to re-check on each run.

Second, the correction lives beside the text, not inside it. You do not rewrite your articles to wedge a phonetic spelling into the middle of a sentence, and you do not regenerate audio already produced so that a new entry applies to future ones. The lexicon is applied at reading time, upstream of the voice.

Third, a word learned once holds forever and everywhere. The day you teach the tool that "Guebwiller" is said "Gweb-vee-lair", every article that contains the name, present and future, benefits. That is the difference between fixing one audio file and teaching a pronunciation.

The property that surprises is deliberate and documented: the corrected word also changes in the on-screen text, not only in the sound. If you use read-along highlighting or subtitles, the pronunciation spelling is what appears at the moment the word is spoken. The reason is mechanical: timestamps have to match what is actually said, or the highlight drifts and marks the wrong word. For most entries this is invisible; for a visible proper noun it is better known in advance than discovered while listening.

Why "add" and not "replace"

An engineering detail with a direct consequence for you: adding a word must never erase the others. A list of pronunciations grows one name at a time, over months, and it represents real editorial work. Wiping it whole because you meant to set a single entry would be the worst kind of accident, all the more so because a missing name goes unnoticed: it raises no error, shows up nowhere, and is only found by listening to a published article weeks later. So WeDispatch merges new entries into the existing lexicon rather than replacing everything, and a word declared twice has its pronunciation updated, never duplicated. The list keeps its order, so that someone re-reading it in their account does not see lines shift on their own.

What it does not do

The lexicon is no magic wand, and saying so is part of our stance. It holds up to two hundred entries per account, which is generous for real use but not infinite: it is a list of special cases, not a full dictionary of French. Each side of an entry is length-bounded, and some characters are refused so that an entry cannot become anything other than a plain word replacement. A refused entry is flagged, so you are not left waiting for a correction that will never come.

Above all, the lexicon fixes the words you teach it, not the ones you overlook. The foundation of French (liaison, numbers, punctuation) is handled upstream, as the French text to speech page explains. The lexicon takes over where that foundation stops: the names you have to know. So the right practice is to build, once, the list of the twenty or thirty words that recur in your content and snag on the ear, then extend it as you go.

It is the same remedy as for acronyms and proper nouns, covered in acronyms in French audio and in pronouncing proper nouns in news audio. And it is exactly the question to put to any provider before committing, the one behind the protocol in how to evaluate a neural voice: not "does your voice read this word correctly?", but "can I force the right reading, and does that instruction hold across all my articles?". A voice that leaves you no hand on its special cases will, sooner or later, leave a mangled name on air.

Read next

• Reading web addresses, emails and handles aloud in French
• The mute e: the French syllable that appears and vanishes
• Adding an audio version to a WooCommerce store