A priori, in fine, etc.: how a French voice should read Latin

Written French borrows constantly from Latin without translating: a priori, in fine, de facto, grosso modo, curriculum vitae. These phrases are so ordinary that people forget they come from elsewhere, until a machine reads them aloud. That is the moment they reveal whether the engine truly reads French text or falls back on the reflexes of another language. And the usual culprit is well known: a model trained mostly on English reads Latin the way an English speaker would, not the way a French speaker does.

French Latin is not English Latin

The point that gets forgotten is that Latin has no single pronunciation. French has its own tradition, shaped by centuries of school and legal usage, and it differs sharply from both the English reading and the ecclesiastical one. In French, "a priori" is said "a pree-o-ree", not the English "ay pry-OR-eye". "Curriculum vitae" ends in "vee-tay" in French, never in the English "VY-tee" or "VEE-tie". "Alea jacta est" is said "a-lay-a yak-ta est", with the j sounded as a French y.

This is exactly the mechanism described in the article on French voices that sound English: a multilingual model has seen far more Latin read by English speakers than by French speakers, and without tuning it applies the majority reading. Latin is where that bias is heard most, because the word is foreign in both languages and the engine has no French spelling cue to hold on to.

| Phrase | Expected reading (French) | Anglicised wrong reading |
|--------|---------------------------|--------------------------|
| a priori | "a pree-o-ree" | "ay pry-OR-eye" |
| in fine | "an fee-nay" | "in fyne" |
| curriculum vitae | "ku-ree-ku-lom vee-tay" | "keu-rikuleum VY-tee" |
| statu quo | "sta-tu ko" | "STAY-toos kwoh" |
| grosso modo | "gro-so mo-do" | "GROH-soh MOH-doh" |
| mea culpa | "may-a kul-pa" | "MEE-a CULL-pa" |
| vice versa | "veece vair-sa" | "VYCE VER-sa" |

The abbreviation that traps everyone: etc.

There is one case where even correct tools stumble, and it is everywhere: "etc." It is not an acronym, it is the abbreviation of "et cetera", and in French it must be read "et say-tay-ra". An engine that spells it out, pronounces it as a single word, or stops on the full stop as if the sentence had ended makes a mistake that leaps out in the very first paragraph. Reading it correctly requires the engine to expand the abbreviation before pronouncing it, exactly as it must expand acronyms and abbreviations, but here without spelling: you do not say the letters, you say the full words.

The same work applies to the other scholarly abbreviations of written French. "Cf." should be read "confer" (or "see"), not "c-f". "N.B." reads "nota bene". "P. ex." reads "par exemple". "Ibid." and "op. cit." belong to the same register. And "vs", very common, reads "versus", not "v-s". None of these forms can be guessed from the letters: the engine has to know them as abbreviations to expand, and know which is said in Latin and which is translated.

What our engine does, without overselling

On the most common Latin phrases in French (a priori, a posteriori, de facto, in fine, in situ, grosso modo, statu quo, mea culpa, curriculum vitae) and on the most frequent abbreviations (etc., cf., N.B., p. ex., vs), the engine applies the expected French reading, expands what needs expanding and avoids the English accent. This is the heart of the stance set out on the French text to speech page: reading one language well, with its borrowings and its conventions, rather than reciting Latin with a foreign accent.

But the limit is worth stating, as with English loanwords. Rare phrases (modus operandi, casus belli, in extenso, ad libitum, sui generis) and the Latin specific to a profession, notably law and medicine, are not all covered. On those, an engine can fall back on the default reading, and nothing in the text stops it on its own.

The move that guarantees the right reading

When a phrase matters, because it recurs across your articles or belongs to your field, the guarantee comes from the pronunciation lexicon. You declare the expected reading once, and the engine stops guessing: a law firm that constantly writes "in solidum" or "intuitu personae" fixes their reading once and never touches it again.

To judge a tool in a minute, give it three phrases: "Il faut trancher a priori" (expected: a pree-o-ree), "Envoyez votre curriculum vitae" (expected: vee-tay) and "des fruits, des légumes, etc." (expected: et say-tay-ra). An engine that clears all three reads Latin like a French speaker. An engine that says "ay pry-OR-eye" or trips on "etc." is running on another language's reflexes, and you will know it before you fit out a whole site.

Read next

• SRT and VTT subtitles: one more file, not one more invoice
• The œ ligature: the little sign that gives a French speech engine away
• Adding an audio version to your Craft CMS articles