French slang, verlan and text speak: what text-to-speech actually does with them
Newsrooms do not write in slang, but their sources do. We ran six real-world French sentences full of verlan, SMS abbreviations and spoken contractions through our reading pipeline. What gets fixed, what is left to the voice engine, and the two settings that handle the rest.
French journalism is written in standard French. The people it quotes are not. A teenager outside an exam hall, a fan after a match, a post lifted from social media: informal French enters an article through quotation marks, and nobody flags it on the way in. On the page it causes no trouble, because every French reader knows what "MDR" means. Read aloud, it becomes a practical question: what does the listener actually hear when the voice reaches that word?
If you publish in French, or you are evaluating a French voice for a client who does, this matters more than it looks. Informal French is not a handful of words. It includes verlan (syllables reversed: "chelou" for "louche", "relou" for "lourd"), borrowed slang ("wesh", "kiffer"), dropped sounds in speech ("j'suis", "y'a") and an entire written shorthand from texting ("tkt", "jsp", "bcp", "2m1").
The test: six sentences, unedited output
Rather than describe what a reading pipeline ought to do, we put six sentences through ours. This is the step that decides how a French text is pronounced before the voice engine receives it. The results:
- "C'est chelou, wesh, on était trop deg", says a student. Nothing changes. Every word goes to the engine as written.
- A user posts: MDR, PTDR, LOL, jsp, stp, bcp. MDR becomes "emme dé erre" and PTDR "pé té dé erre", the French letter names. LOL, jsp, stp and bcp are left alone.
- "J'suis pas venu, y'a eu un souci." Nothing changes. Spoken contractions pass through.
- A sentence of pure slang ("ouf", "dinguerie", "meuf", "kiffe grave"). Nothing changes.
- He said OKLM, then IRL, then XPTDR. XPTDR is spelled out letter by letter. OKLM and IRL are left to the engine.
- She wrote tkt, a+ and 2m1. Nothing changes, not even the numeral in "2m1" (texting for "demain").
The pipeline knows around forty families of written forms that have a correct spoken reading in French: numbers, dates, units, legal references, phone numbers, acronyms. Slang is not one of them, by design. There is no rule for it. "Ouf" is read the way it is spelled; "bcp" cannot be read at all.
Why capitals get spelled and lowercase does not
The only rule that touches any of these words is the acronym rule, and it does not try to understand the word. It checks whether the letters can be pronounced as a syllable. No vowel, or three consonants at the start, means the group cannot be said as a word, so it is spelled out. That is why MDR comes out as letter names, which is what a human reader would do too. The same test hands LOL, OKLM and IRL to the engine because a syllable is possible. For LOL that is fine: French speakers say it as a word. For OKLM ("au calme") and IRL, a spoken word is likely the wrong outcome. We cover the acronym rule and where it stops in acronyms in French text-to-speech.
Lowercase abbreviations go through no rule at all. "stp", "jsp" and "tkt" reach the engine untouched, and what happens next is not consistent from one word to another: spelled, half-pronounced, or swallowed. We would rather say so than claim a behaviour we do not control. The voice is tuned to read French well, and the French it reads well is French written out in full.
What will not be flagged for you
Before an article is read, we look for words whose pronunciation is uncertain and raise them for a decision. The detection relies on one dependable signal: a capitalised word in the middle of a sentence is probably a proper noun. That is how it finds footballers' surnames and Breton villages without a dictionary, as explained in names flagged before reading.
For slang, the consequence is mechanical, and we checked it on the same six sentences. OKLM, IRL, PTDR, XPTDR and LOL can come up as candidates, because they are in capitals. "chelou", "wesh", "tkt" and "bcp" never do, because they are lowercase. The only safety net for those is a human ear.
Two settings that take care of the rest
The lexicon, for what keeps coming back. A newsroom that quotes social media sees the same abbreviations again and again. One entry per word is enough: "bcp" read as "beaucoup", "OKLM" read as "au calme". An entry matches the whole word, ignores case, and applies to every later article. It also runs before the French reading rules, so an entry for MDR ("mort de rire") wins over spelling it out. What readers see does not change: the on-screen text, the read-along highlighting and the subtitles keep "bcp"; only the engine hears "beaucoup". The steps are in fixing how a word is pronounced. One limit worth planning for: the lexicon holds two hundred entries, and local place names usually deserve them more than an abbreviation seen once.
The read text, for the one-off. For a single quote, the audio's detail page lets you replace the text that is read without touching the published article. You write "tu sais pas" where the source wrote "jsp", and the article keeps the exact quotation. It is the general principle of what is said versus what is shown, applied to the case where faithfulness on the page and clarity in the ear pull in different directions.
An editorial call, not a technical one
One decision stays with you: should "MDR" be spelled or expanded? Spelling respects the source; expanding helps a listener who has never seen the abbreviation. A youth outlet and a regional daily will answer differently, and both can be right. What matters is deciding once, recording it in the lexicon, and not revisiting it every morning.
For verlan and slang written in full, pronunciation is rarely the problem. The engine reads what is written, and what is written is how it sounds. The risk lies in tone: a casual quote delivered in the even voice of a council report can sound odder than the word itself. If quote voices are switched on, a full statement in quotation marks that opens a sentence is read by a different voice, which marks the change of speaker. A short phrase inside a sentence stays in the article's voice, deliberately; the reasoning is in quotes read by a second voice.
Give your articles a voice with WeDispatch
This blog is itself voiced by WeDispatch. Curious how it sounds on your content?
Book a demo