The diaeresis out loud: how French text to speech handles Noël, maïs and aiguë
The French tréma tells the ear that two vowels are pronounced apart. When it goes missing, a speech engine reads maïs as mais and Noël as one syllable. What a French engine does, and how to fix it.
There is a tiny mark in French, two dots over a vowel, that alone decides whether a word is read as one syllable or two. The diaeresis (in French, the tréma) does not change the quality of a letter the way an acute accent does: it separates. It says the vowel carrying it does not blend into the one before it, but is sounded on its own. "Mais" and "maïs" differ only by those two dots, yet one is a single syllable (the conjunction but) and the other is two ("ma-eess", the cereal maize). For a speech engine, this is a quiet and unforgiving test, because the mark is fragile and its loss leaves no visible trace in the text.
What the diaeresis asks of the ear
Take five common French words and listen for the break the diaeresis forces. The right-hand column is the error you hear when the engine treats the word as if the mark were not there.
| Written word | Expected reading | Common error |
|--------------|------------------|--------------|
| maïs | ma-eess (two syllables) | "meh", mistaken for the conjunction mais |
| Noël | no-el (two syllables) | one syllable, "nurl" |
| aiguë | ai-gu, the u is sounded | "aig", the u swallowed as in langue |
| coïncidence | co-in-ci-dence | "kwan", the two vowels fused |
| canoë | ca-no-eh | read as English canoe |
What these five cases share is that none can be guessed from the letters alone once the mark is ignored. Remove the two dots and the word becomes a different word, or a non-word that the engine snaps to the nearest thing it knows. "Aiguë" without its diaeresis would rhyme with "figue"; with it, the u finds its voice again. This is the kind of finish that separates a careful reading of French from letter-by-letter decoding, and it is the heart of the work described on the French text to speech page: how a word is pronounced depends on marks the writing carries discreetly, and that an engine has to read for real.
Why the diaeresis so often disappears
The real danger is not that the engine ignores the diaeresis when it is present. Good models respect it. The danger is that the mark has already vanished from the text before it reaches the voice, and for very ordinary reasons.
On a keyboard, the diaeresis is composed in two steps, and many people never type it. A capital letter often loses its marks, the same trap covered for the accented capitals: a headline set as "NOEL" has no diaeresis left to read. A paste from an old application, a PDF extraction, a badly encoded database: at each step, two small dots can drop. And the day "maïs" becomes "mais" in the source, the engine makes no mistake, it faithfully reads what it was given: the conjunction. The fault is upstream, but it is the ear that pays for it.
Proper nouns, where nothing is forgiven
The diaeresis is common in first names and brand names: Loïc, Anaïs, Gaël, Raphaël, Israël, and above all Citroën, whose diaeresis on the e separates the o from the e ("sitro-en") and which many engines read "sitrun" as though it were an English word. On a misread first name in a bulletin, a place name, a brand cited in an article, the error is heard at once and signals that the audio was bolted on without proofreading.
For those words, the guarantee comes not from the voice but from the pronunciation lexicon. You declare once how "Citroën" or "Anaïs" must be read in your content, and the correction holds across every later article, whether the diaeresis survives in the source or not. It is the same principle as for proper nouns and place names: spelling alone does not always give the expected sound, and the lexicon is where you settle it, leaving nothing to chance.
Testing a tool in one minute
Three words are enough to know where you stand. Have the engine read "le maïs pousse", "joyeux Noël" and "une douleur aiguë". A tool that says "ma-eess", "no-el" and "ai-gu" genuinely reads the diaeresis. A tool that hands you "meh", a single syllable and a mute u treats the mark as noise, and you will know you have dozens of words to fix by hand across a whole site. It is the same reflex as for reading numbers aloud: you do not judge an engine on a nice paragraph, you judge it on its traps.
The diaeresis, the cedilla and the rest
The diaeresis is the most fragile mark because it is the easiest to lose, but it is not alone. The cedilla on the "ç" (français, garçon, reçu) is generally better handled, because it is harder to drop without making the word visibly wrong. The 1990 French spelling reform also moved the diaeresis in "aiguë" onto the u, giving "aigüe": both spellings exist, and a serious engine should read them the same way.
The lesson is simple. A diacritical mark is not decoration: it is a reading instruction. Respecting it when it is there, and knowing how to restore it when it has gone, is exactly what separates a tool that pronounces French from one that recites it from memory.
Give your articles a voice with WeDispatch
This blog is itself voiced by WeDispatch. Curious how it sounds on your content?
Book a demo