Reading legal references aloud in French: what a voice tool does with « L. 121-1 » and « al. 2 »

A court report, a legal note, an article commenting on a ruling: these texts are built from references that fit in a few characters and take a completely different shape when spoken. In French, « L. 121-1 » is said « L cent vingt et un », « al. 2 » is said « alinéa deux », « 49.3 » is said « quarante-neuf trois ». A synthetic voice that does not know these conventions reads the characters in the order it receives them, and the result gives the machine away at once: a hyphen spoken as a hyphen in the wrong place, two numbers stuck together read as a football score, an abbreviation swallowed whole. So the honest question, before you hand a legal text to a voice tool, is this: what exactly does it do with those references?

Why this is a language problem, not a volume one

The trouble is not that these forms are rare. It is that they are ambiguous. « 222-13 » can be an article of the penal code, a page range, a serial number or a score. « 2019/882 » can be a regulatory reference, a fraction or a sporting season. A model that guesses gets it wrong one time in ten, and a reference wrongly transformed is worse than one left raw: the second is still audible, the first misleads. This is the same stance that guides our work on French text to speech: when in doubt, leave it alone. Every rule below is deliberately narrow, triggered only by a reliable cue.

Five real references, and how they are read

Here is what our text preparation produces today, measured on real sentences.

« Selon l'article L. 123-4 du code de la consommation » comes out « article L cent vingt-trois tiret quatre ». The letter is left as is (the engine pronounces it correctly), the number is spelled out, and the hyphen is named, because a lawyer names it. Three-part forms follow the same logic: « articles L. 122-5-1 » gives « L cent vingt-deux tiret cinq tiret un ».

« L'article 222-13 du code pénal » comes out « article deux cent vingt-deux tiret treize ». Half the references in a court report carry no letter, and they used to pass through untouched. One condition matters: the word « article » or « alinéa » must sit before the number. Without it, a bare « 222-13 » stays raw, because nothing proves it is a reference rather than a score. That is a deliberate abstention, not an oversight.

« Le gouvernement a recouru au 49.3 » reads « au quarante-neuf trois », and « l'art. 49.3 de la Constitution » gives « l'article quarante-neuf trois »: the abbreviation « art. » is expanded along the way. In front of a unit, though (« une hausse de 49.3 % »), the same form stays a decimal: « quarante-neuf virgule trois pour cent ». Context decides.

« La directive 2019/882 du Parlement européen » reads « directive deux mille dix-neuf, huit cent quatre-vingt-deux »: the slash becomes a pause, and the number is spoken in full, the way a human reader would, not digit by digit.

Where it stops, said plainly

The fifth form is where the tool shows its limit. « Voir art. 50 et § 4 » reads « voir article cinquante et § quatre »: the section sign is not expanded. It stays in the text as it is, and the engine makes of it what it can, which is usually nothing good. This is a real shortcoming, and it mostly affects very technical texts (codes, internal regulations) where « § » is common. We would rather say so than hide it: a page that promises to read the law aloud has to own what it does not read yet.

The same honesty applies to agreement: « 21 articles » read aloud stays « vingt et un articles », because matching the number to the gender of the noun that follows would need a dictionary the reading does not carry. The flaw is old, it is known, and it is not fixed where we do not write the number ourselves.

The fallback exists, and it is in your hands

For anything still said wrong (a « § », a reference without its trigger word, a Latin term specific to your field), the pronunciation lexicon is the right answer. You correct it once, and the product remembers it for the whole newsroom, exactly as you would for a proper noun: our piece on fixing word pronunciation walks through the gesture. And for acronyms, which swarm in a legal text, the rule is set out in our page on acronyms and abbreviations.

At bottom, reading the law aloud is the textbook case of a job where language matters more than quantity. Law firms and legal desks that voice their texts know this, and it is for them that we work on these rules, as we explain about legal blogs and law firms. A reference said well goes unnoticed; a reference said wrong stops the listening. That is the whole difference.

Read next

• Listening to a long text in French TTS: listening fatigue, and what actually helps
• Choosing what the audio reads: skip the boxes, the « read also » blocks and the rest
• Listening to an article with the screen locked, and in the car