Août, but, persil: the French words native speakers pronounce differently, and who decides for the voice
A handful of everyday French words have two accepted pronunciations, and a few have a rule that usage ignores. A synthetic voice still has to pick one. What the norms say, why a text-to-speech tool should not settle the question for you, and how to impose your house choice without breaking other words.
If you publish in French, or audit French audio for a client, here is a small surprise: some ordinary words do not have one correct pronunciation. Ask five native speakers to read "le 15 août", "marquer un but" and "du persil" and you will likely hear both "oo" and "oot" for août, both "bu" and "bute" for but, both "persee" and "perseel" for persil. Nobody is wrong. Standard dictionaries record both forms.
A human reader resolves this without thinking, according to region, age and habit. A synthetic voice cannot avoid the choice: every occurrence gets one reading or the other, and on a news site that choice repeats across hundreds of articles. So the useful question is not only which form is right, but who gets to decide.
Three kinds of disagreement
It helps to separate the cases, because they call for different responses.
Two accepted pronunciations. These are genuine variants, and neither is a mistake.
- août (August): "oo" or "oot". An older "a-oo" survives in places.
- but (goal): with or without the final t. The t is now dominant among younger speakers and in sports commentary, which matters if you publish match reports.
- persil (parsley) and nombril (navel): the final l is silent in the older form, sounded in the newer one.
- mœurs (customs) and ananas (pineapple): final s optional.
A norm that usage ignores. The rule is clear, and many speakers do not follow it.
- gageure (a challenge) is "ga-zhure": the e only softens the g. "Ga-zheur" is widespread but considered incorrect.
- dompter (to tame) and sculpter: the p is traditionally silent, though sounding it is spreading.
- second is pronounced with a g, despite the c.
One spelling, two meanings, two pronunciations. This is the dangerous group, because the choice is not taste but sense.
- Cassis: the blackcurrant sounds its s, the town on the Mediterranean coast does not.
- fait: "un fait" (a fact) may take a t; "il fait beau" (the weather is fine) never does.
- plus, tous, fils: the reading depends on the sentence. We cover these in French homographs in text-to-speech.
Why the tool should not choose for you
Faced with two accepted forms, a speech model reads the one it heard most during training. That is neither an error nor an editorial decision. It is an average. An average may suit a national outlet and sound slightly off to an audience in eastern France or Belgium, where "vingt" (twenty) is often said with a t that Paris drops.
That is why our shared pronunciation layer, the one that would apply to every customer without being asked, contains none of these words. In fact it is empty for now, on purpose. A word can only enter it once it has been heard and confirmed, and once the reading holds for two newsrooms that have nothing to do with each other. A word whose pronunciation is a house preference fails that second test by definition. It stays with you. That restraint, correct only what has been heard and never impose an average, runs through our whole approach to French text-to-speech.
We are also not publishing a word-by-word table of what our voices do with each example. The result depends on the voice and the sentence, and such a table would go stale at the next model update. It is quicker and more reliable to listen yourself, using the method in test a French voice in ten minutes: drop three or four of these words into a typical sentence from your own articles.
How to impose a house choice
If the default reading is not the one you want, the pronunciation lexicon handles it. One line per word, in the form "word = pronunciation", written in ordinary French spelling rather than phonetic symbols: "août = oute", "gageure = gajure".
Four properties of the lexicon matter for this use, all of them taken from how the code behaves:
- Whole words only. An entry for "but" leaves "butte" and "début" alone. Longer entries are applied before shorter ones, so a compound name is never broken by one of its parts.
- Every occurrence, regardless of case. "Août" at the start of a sentence is corrected like "août", and the capital letter follows your text.
- Before every synthesis, including in Faithful reading mode. It is a deterministic replacement applied before the voice receives the text, not a hint the model may ignore.
- Up to two hundred entries per account. Plenty for a house list, not meant to be a dictionary.
The first two properties are exactly why the third group must stay out of the lexicon. Because an entry applies everywhere the word appears, "fait = fète" would turn "il fait beau" into nonsense, and "Cassis = Cassi" would rename every glass of crème de cassis after a seaside town. Words whose pronunciation depends on meaning are better left to the engine, which reads context better than any replacement rule.
The short version
For a word with two accepted forms, choose once, put it in the lexicon, and stick to it: listeners notice inconsistency far more than they notice which variant you picked. For a word with a clear norm, follow it if your audience cares. For a word whose sound depends on its meaning, do not override anything until you have listened.
Give your articles a voice with WeDispatch
This blog is itself voiced by WeDispatch. Curious how it sounds on your content?
Book a demo