Proofreading French by ear: what text to speech catches, and what it never will

Having a synthetic voice read your draft back to you is one of the most effective proofreading tricks around, and plenty of English-speaking writers swear by it. The reason it works is well known: your eyes read what you meant to write, your ears hear what is actually on the page. A synthetic voice has no idea what you meant. It just reads.

If you write in French, or edit French copy as a second language, the trick still works, but less completely than you might assume. French spelling carries a large amount of grammar that is never pronounced. On top of that, the tool doing the reading may have adjusted your text before speaking it. Here is what listening reliably catches, what it structurally cannot, and how to build a routine that accounts for both.

What the ear is good at

Anything that breaks rhythm or syntax jumps out when heard, even when the eye glides over it:

• a missing word: "il a décidé reporter la séance" (he decided postpone the meeting) looks fine at a glance and snags immediately when spoken;
• a doubled word, especially across a line break: "le le conseil";
• a half-edited sentence, where you rewrote the opening and forgot to adjust the ending;
• a sentence that runs too long: by the time the voice reaches the main verb, both you and the listener have lost the subject;
• repetition: three uses of "important" in one paragraph, invisible on screen, grating aloud;
• an acronym never spelled out, because you know it so well you stopped seeing it.

None of these are spelling mistakes. They are writing mistakes, and no spell checker flags them. Listening is the best tool there is for them, which is also the starting point of our guide on writing for the ear.

What the ear cannot hear: silent grammar

English has its homophones (their, there, they're), but French has whole layers of grammar that disappear when spoken. These are features of the language, not limits of any particular voice, and no reader, human or synthetic, can make them audible. Each pair below sounds identical:

• past participle agreement: "les mesures qu'il a décidé" and "qu'il a décidées" (only a handful of participles ending in a consonant, like "prises", change sound in the feminine);
• -é, -er and -ez endings: "il faut décidé" and "il faut décider" are pronounced the same;
• -ait and -aient endings: "il votait" and "ils votaient"; so "les élus votait", which is wrong, sounds right;
• silent plurals: "les adjoint" sounds like "les adjoints"; "ils parle" like "ils parlent";
• grammatical homophones: ces / ses / c'est / s'est, son / sont, ce / se, quel / quelle / qu'elle, a / à, ou / où.

There is one partial exception: liaison. "De grands espoirs" is said with a [z], "de grand espoirs" with a [t], so a voice that makes the liaison betrays the missing agreement. But optional liaison is exactly where machine reading of French is least predictable, as we explain in French liaison and text to speech. You cannot rely on it as a proofreading signal.

The practical upshot: for French, listening complements a grammar checker rather than replacing it. The two catch different errors with very little overlap. An English writer who has learned to trust the ear alone will miss more in French than they are used to.

The less obvious trap: the tool may fix things first

A text to speech tool rarely reads your raw text. Between the page and the voice there is a preparation step, and it can change what it receives. For proofreading, that matters: if the tool repaired a mistake before speaking, the audio sounds right while the page stays wrong.

WeDispatch has two preparation routes, and both can do this.

In standard mode, a language model prepares the text for reading. It is instructed not to rewrite, but a model can smooth an awkward phrase or restore a dropped word without saying so. The audio may come out cleaner than your copy.

In the mode where no model touches the text, preparation applies only written rules. One of them restores a closed list of commonly dropped French accents and cedillas: "ca" becomes "ça", "tres" becomes "très", "etait" becomes "était", "deja" becomes "déjà". The list is deliberately cautious. It never touches a form that is itself a valid French word (a and à, ou and où, sur and sûr, cote and côte are left alone), and it skips words written entirely in capitals, so that "CA", the usual abbreviation for revenue, does not turn into "ÇA". When one of these corrections is made, the preparation log records it as a line saying the accents and cedillas were restored.

That is the right behaviour for listeners, who should never hear "ka va". It does mean that if you use the audio to proofread, a missing accent from that list will not be audible. The same applies to the French text to speech tools you might compare us with: ask each one what it does to your text before it reads it, and whether it tells you.

A proofreading routine that accounts for all this

Three passes beat one:

1. Grammar checker first, for the silent grammar: agreement, verb endings, homophones. That is its territory, not the ear's.
2. Listen second, without reading along, or following without stopping. Mark every point where you lose the thread or have to go back. Almost every one of them is a sentence to rewrite.
3. A final look at what the tool might have quietly fixed: dropped accents from the list above, split words, and any passage you knew was shaky that sounded surprisingly smooth.

For more on what happens when accents go missing altogether, see French text typed without accents. Listening is an excellent test of how well a text is written. It is not a spell checker, and in French it is further from being one than in English.

Read next

• Security questionnaire for an audio tool: what to ask, and our answers, including the noes
• Breton, Basque, Corsican, Alsatian: why French text-to-speech trips on regional names
• A woman's or a man's voice for a quote: what French grammar reveals about who is speaking