French TTS and medical vocabulary: where a French voice stumbles, and how to fix it

If you publish health content in French (a hospital in Quebec, a Swiss health insurer, an international NGO with French-speaking programmes, a pharma company's patient pages for France and Belgium), there is a good chance the person checking the audio version is not a native speaker. That matters, because medical French is where a synthetic voice is most likely to go wrong, and where the errors are hardest to spot unless you know what to listen for.

The stakes are also higher than in general news. A mispronounced place name is a curiosity. A mispronounced clinical term makes the listener doubt the whole page, and the people who listen most on health sites are patients, carers, older readers and people with low vision. Here is a practical map of the traps, and what you can do about each one.

Five families of traps

"ch" that should sound like "k"

In French, "ch" is normally the sound in "shoe". In many words of Greek origin it is a "k" instead: cholestérol, chlore (chlorine), ischémie (ischaemia), chorée (chorea), chondrite, orchite. English speakers find this intuitive, since English does the same. The catch is that French is not consistent: chirurgie (surgery) and chimiothérapie (chemotherapy) use the "sh" sound. A voice gets the common words right and guesses at the rest.

Final consonants that are spoken

Most final consonants in everyday French are silent. Medical vocabulary breaks the rule: sinus, virus, tétanos, psoriasis and pancréas all sound their final "s". Latin words ending in "-um" are said "ome": sternum, duodénum, septum. And os (bone) sounds its "s" in the singular, and not in the plural.

Ligatures

Œdème, œsophage and fœtus are read with an "é" sound. When the ligature is typed as two letters, "oe", which happens a lot in templates and in text exported from other tools, the risk of a wrong reading goes up.

Acronyms: spelled out or read as words

IRM (MRI), AVC (stroke), CHU (university hospital) and ARS (regional health agency) are spelled letter by letter. SAMU and EHPAD (care home) are read as words. HAS, the national health authority, is spelled out, but an engine may be tempted to read it as a word. Our article on acronyms in French TTS covers the general rules; health content adds dozens more that writers rarely expand.

Prescription shorthand

This is where meaning is at stake. "2,5 mg" needs to become "two point five milligrams" in French (note the decimal comma), "1 cp" is "one tablet", "37,8 °C" has to be read in full. And IV is the textbook case: intravenous on a care sheet, the number four in "Henri IV" or "chapitre IV". No automatic rule can settle that safely.

What gets flagged automatically, and what does not

Before reading, WeDispatch spots proper names whose pronunciation is uncertain and asks you about them, as described in flagging uncertain names before reading. That helps with brand names of medicines, which are capitalised.

Its limit is worth understanding, because it comes from how it works. Detection relies on one rule: a word that keeps its capital letter in the middle of a sentence is almost always a proper name. That is what lets it work without a dictionary. It also means a lowercase word (ischémie, amoxicilline, chorée) is never a candidate. To the detector, it is an ordinary word. Clinical vocabulary, written in lowercase, is exactly the part you have to listen to yourself.

The fix: the pronunciation lexicon

Once you hear a problem word, you correct it in the account's pronunciation lexicon (a feature of the plans that include it). You enter the word as it appears in your copy, and a reading spelling in plain French letters, not in phonetic symbols:

• ischémie → iskémie
• chorée → korée
• chondrite → kondrite
• BPCO → B P C O

The behaviour is deterministic, and we checked these rules against the code rather than the marketing copy:

• whole words only: an entry for "os" leaves "osseux" and "dose" alone;
• case-insensitive, accent-sensitive: a capitalised "Ischémie" at the start of a sentence is corrected too, and keeps its capital;
• longer entries first: a multi-word entry wins over a shorter one it contains;
• the displayed text does not change: the voice receives "iskémie", while the read-along text and subtitles still show "ischémie";
• two hundred entries per account, and anything beyond that is refused with a message, never silently dropped.

Case-insensitivity has one direct consequence. An entry "IV → intraveineux" would also rewrite "Henri IV". For context-dependent abbreviations, use a multi-word entry ("en IV → en intraveineuse") or, simpler, write the term out in the article. The lexicon is meant for words whose reading never depends on context.

What we do not claim

We do not publish a list of medical terms our reading gets right or wrong. It would depend on the voice, and it would be out of date at the next engine update. Nor do we ship a shared medical dictionary for every account: the shared pronunciation base is empty today, on purpose, because a wrong entry there would apply to every customer at once. Nothing goes into it before someone has listened to it.

So the honest method is yours, and it is short. Take the twenty terms your site uses most, have them read inside real sentences, note the ones that snag, and add them to the lexicon. If you do not speak French fluently, ask a French-speaking colleague to listen to that one list rather than to whole articles. Our ten-minute test protocol works as is for this, and the French text to speech page explains the text preparation that happens before the voice reads anything.

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