French isn't read the way it's written: what that asks of a synthetic voice

Some languages are read roughly the way they are spelled: one letter, one sound, done. French is not one of them. Between what sits on the page and what a voice has to produce, there is a constant gap, and closing that gap is the actual job of a text-to-speech system. A machine reading French letter by letter would be unlistenable. Everything interesting happens in what it chooses not to say, to say differently, or to add.

Here are the four places where that gap opens up, and what an honest tool does about each.

Letters you write and don't say

French is full of silent letters. Most final consonants go unpronounced: the "t" in chat, the "s" in temps, the "d" in grand, the "x" in deux. Verb endings are riddled with them: ils parlent sounds exactly like il parle, with the "-ent" silent. A human reader sorts this out without thinking, because they recognise the word before they say it. A synthetic voice has to have learned that those letters are there for the eye, not the ear.

The trap is that the rule has exceptions that are pronounced: the "s" in fils, the "t" in but, the "c" in sac. A model trained on a lot of French handles the common cases; where it hesitates is on rare words and, above all, on proper nouns, which follow no general rule at all.

Liaison, which adds a sound that isn't written

The reverse also happens: French sometimes conjures a sound that the isolated word doesn't contain. Les enfants: the "s" of les, silent before a consonant, wakes up as a "z" before the vowel. This is liaison, and it is probably the hardest thing for a machine, because it is neither always required, nor always forbidden, nor always optional. Saying un grand-z-homme is wrong, saying les-z-amis is mandatory, and between the two lies a grey zone where even native speakers disagree. We gave it a whole article, because it tells you so much about a voice: liaison, the hardest test for French text-to-speech.

The e you keep or let drop

The unaccented "e", the so-called mute or schwa "e", is another running negotiation. In samedi you barely hear it; in une petite fenêtre it vanishes almost everywhere in speech, yet would return in careful diction or in verse. Pronounce every "e" and the reading sounds stiff and foreign; drop them all and it turns to mush. The right setting depends on the register of the text and the pace, and it is a reading decision as much as a rule. What our engine does with it is here: the mute e.

Words spelled alike, said differently

Finally, some words that look identical change their pronunciation with meaning or grammatical role. Le président (the president) and ils président (they preside) are not said the same way; est the verb and est the compass direction part company out loud. These are homographs, and they force the voice to understand a little of the sentence before reading it, since spelling alone does not decide. It is ground where a machine still slips, and where it should sometimes hold back rather than gamble. The detail is here: homographs.

What a tool can do with the gap

The first thing to expect from a French text-to-speech tool is not a beautiful voice, it is a reading that has resolved these four gaps correctly across ordinary text. A model tuned for French handles silent consonants, common liaisons and dropped "e" well; it still stumbles, as it should, on proper nouns, rare words and homographs whose sense isn't obvious.

The second thing to expect is a way to fix what remains. No model guesses that a particular surname should be said a certain way: you need to be able to tell it once, and have it remember. That is the point of a pronunciation lexicon, which turns a one-off error into a durable setting instead of letting it come back on every article.

What to refuse is the tool that claims to have none of these problems. French is not read the way it is written: a voice that never erred on a proper noun does not exist, and the one that promises it is lying about the rest. The right question is not "is it perfect", it is "is it correct everywhere the text is ordinary, and fixable everywhere it isn't".

Hear it on your own text

The best way to judge an engine on this gap between writing and speech is to hand it your own text, with your own names and vocabulary. The free tier lets you do exactly that, with no commitment: try it now.

Read next

• How many credits one audio costs, and how to know before you make it
• Add the audio of your articles to a beehiiv newsletter
• French text to speech: the silences that let a long article breathe