Roman numerals in French audio: Louis XIV is not read like the XIVe siècle

A Roman numeral hides both its value and its grammatical form. "XIV" is fourteen, but it reads "quatorze" after a king's name and "quatorzième" before the word "siècle". "Ier" reads "premier", never "one". And "II" reads "deux" in "tome II", "deuxième" in "la IIe République". Where a human reader decides without thinking, guided by the sense of the sentence, a speech engine has to infer, and this is one of the fastest ways it gives itself away on carefully written French.

For anyone publishing history, law, culture or institutional content, this is not a minor detail. Sovereign names, centuries, republics, dynasties, volumes and legal articles are full of Roman numerals. A voice that spells out "X I V" or says "Louis le quatorzième" instead of "Louis quatorze" does more than fumble: it tells the ear that no one checked the reading, on exactly the content where precision is the point.

The rule, in one line

A Roman numeral is read as a cardinal when it numbers a person within a line (king, pope, emperor) or a unit of a set (volume, chapter, article, act). It is read as an ordinal when it marks a rank in a series (century, republic, dynasty, district, olympiad). The one constant exception is "Ier", always read "premier", whatever the context.

The table of forms that fool the machine

| Written form | Expected reading | Common error |
|---|---|---|
| Louis XIV | Louis quatorze | "Louis le quatorzième", "Louis X I V" |
| François Ier | François premier | "François un er", "François un" |
| Napoléon III | Napoléon trois | "Napoléon troisième" |
| le XIXe siècle | le dix-neuvième siècle | "le X I X e siècle", "le dix-neuf siècle" |
| la Ve République | la cinquième République | "la V e République", "la cinq République" |
| tome II, chapitre IV | tome deux, chapitre quatre | "tome deuxième", "tome I I" |
| l'article VII | l'article sept | "l'article septième" |
| Benoît XVI | Benoît seize | "Benoît seizième" |
| la Seconde Guerre (39-45) | la Seconde Guerre mondiale | (reverse trap: do not romanise) |
| MMXXVI | deux mille vingt-six | spelled letter by letter |

Each line is a two-minute test you can run against any tool, ours included. The right-hand column is not theory: these are the readings you hear on voices that recognise the run of capitals but not the rule that governs it.

Why it is hard for a machine

Two difficulties stack up. The first is recognition: "XIV" looks like a word in capitals, "V" like a plain letter, "Ier" like a clipped word. An engine without a dedicated rule either spells or invents. The second, deeper one is the cardinal-versus-ordinal decision, which depends on the neighbouring word and sometimes on the whole meaning. "Louis XIV" and "le XIVe siècle" carry the same numeral and call for opposite readings. No local phonetic prediction is enough: the engine has to look at what the numeral is counting.

Then there is the reverse trap, over-recognition. Not every run of capitals that looks like a Roman numeral is one. "Vitamine D" is not five hundred, "la ligne C" stays line C, an initialism that happens to spell "MI" is not "one thousand and one". A greedy rule damages the text instead of helping it, exactly the failure we work to avoid everywhere in the pipeline.

Good preparation therefore means normalising upstream: spotting a Roman numeral, deciding from its immediate context whether to read it as cardinal or ordinal, and touching it only then. It is rule work, not timbre, and it is the underlying logic of the French text-to-speech page: the correct reading is prepared before synthesis.

What WeDispatch does, without overpromising

The regular cases in the table, sovereigns and popes as cardinals, centuries and republics as ordinals, volumes and articles as cardinals, "Ier" always premier, are handled by the normalisation applied to every article. That is the baseline.

What remains is what context alone cannot settle. A lone "III" in a heading, with no word to say whether it is a volume or a dynasty, keeps a share of ambiguity that no rule removes honestly. There, two things. First, a well-written heading helps the ear as much as the eye: "Partie III" reads without hesitation, "III" on its own does not. Second, for a reading that must hold across a whole corpus, the pronunciation lexicon lets you fix the expected form once, with no need to regenerate articles already online.

We would rather put it that way than claim perfect guessing. The machine applies solid rules across the great majority of occurrences, and leaves you in control of the rare cases where meaning alone decides.

The test to run before you choose

Take a real page that contains Roman numerals, a biography, a historical note, a legal text. Find two or three of different kinds: a sovereign, a century, a volume. Have the tool you are evaluating read them, and listen only to those points. Within a minute you will know whether the tool knows the rule or merely spells.

Roman numerals round out two other revealers of machine-read French: numbers read aloud and dates read aloud. And for an overall method to choose between two voices on your own text, see how to evaluate a neural voice.

Read next

• Adding an audio version to your Magento pages
• Reading French dates aloud: where text-to-speech quietly gets it wrong
• The middle dot out loud: what text-to-speech does with French inclusive writing