← Blog · 20 November 2026 · Lire en français

Auxerre, Craon, Suresnes: French town names that trip up text-to-speech

French place names do not follow French reading rules, and no rule set catches them all: Doullens ends in \"an\", Saint-Gaudens in \"ince\". Examples from our pronunciation bank, why homographs make town names harder still, and a method local newsrooms can apply in an afternoon.

Listeners forgive a synthetic voice a great deal: a slightly even pace, a cautious intonation, breathing that is a little too regular. What they do not forgive is getting the name of their town wrong. Saying "Ock-serr" for Auxerre, or sounding every letter of the Doubs, tells a local audience in two syllables that nobody from around here read this text. For a local news outlet, whose whole value is being close to its readers, that is the most expensive mistake there is.

It is also a very predictable one. A voice tuned to read French properly applies French reading rules. French town names were fixed centuries ago, often in spellings that no longer match how they are said, and every region kept its own habits. The voice does exactly what it was taught, and that is precisely why it gets them wrong.

Why no rule is enough

It is tempting to fix this with rules: "x sounds like ss", "an inner s is silent". The list of French communes quickly shows that this does not hold. A few pairs, all real:

  • Final "-ens". Doullens and Confolens end in a nasal "an". Saint-Gaudens and Val Thorens end in "ince". Same spelling, two readings, and nothing in the letters to tell them apart.
  • "-ouen". Écouen is said "ay-koo-an". Saint-Ouen is said "san-twan". Two towns a few kilometres apart near Paris.
  • Final s. Silent in Thouars, Flers, Ancenis and Tournus. Pronounced in Senlis, Bazas, Jouy-en-Josas and the Gers.
  • Final x and z. Silent in Savoyard names (Chamonix, Oyonnax, Culoz, La Clusaz), pronounced in Alpe d'Huez.

Any rule that fixes one breaks the other. The only method that holds is the humble one: deal with names one at a time, and decide for each.

What our pronunciation bank shows

Since late September we have been preparing, theme by theme, names a French voice usually mangles. The "Towns of France" theme has fifty-two entries, and a neighbouring "Regional names" theme has seventy more (Alsace, the North, the Basque Country, Provence, Corsica, overseas territories). Each entry holds the name as written, a reading spelled in ordinary French letters (never phonetic notation, which the engine does not read) and the reason for the fix. A few examples as they appear in the bank, with the French respelling in quotes:

  • Auxerre → "Ausserre", Auxonne → "Aussonne": the x is said as ss.
  • Craon → "Cran", Raon-l'Étape → "Ran": "aon" is a plain nasal "an".
  • Suresnes → "Su-rène", Fresnes → "Frène", Asnières → "A-nière": the inner s is silent, as in Aisne ("Ène").
  • Millau → "Mi-yô": the double l sounds like y. Yet Vizille → "Vi-zile" keeps a plain l.
  • Longwy → "Lon-oui": silent g, and "wy" sounds like "wee".
  • Gérardmer → "Gé-rar-mé": neither the d nor the final r.
  • Sainte-Menehould → "Sainte-Me-nou": final "-ould" is just "oo".
  • Troyes → "Troie", Bourg-en-Bresse → "Bourk-en-Bresse".

The pattern is useful in itself: most errors come from a handful of French rules (the nasal vowel, the silent final consonant, the inner s) applied where local usage went another way.

What the bank does not do

One clarification matters more than the examples: an entry in the bank does not apply on its own. It is a proposal. It only enters the shared base, the one that applies to every account, after someone has listened to it as is, then corrected, and approved it. The reason is written in our code: a wrong shared base is worse than none. And nothing guarantees that a listed name is actually misread by the voice. The bank records the expected error; only listening tells you whether it happens.

Town names add a specific trap: many of them are also ordinary French words. Sens ("sense"), Guise ("manner"), Tours ("turns"), Bar ("bar"), Vire ("veers", from the verb). A fix applied to "Sens" would change every sentence containing "dans tous les sens", because lexicon matching ignores capitalisation. So the bank rejects, by design, any homograph of an everyday word, and our checks verify this against a corpus of ordinary French. Those names have to be handled case by case, in the text itself or in an account lexicon whose owner knows exactly what it affects.

Finally, some names divide their own residents. Many people in Savoy say "Chamoni", most of France says "Chamonix". There is no universal right answer there, only the one your readers expect.

A method for local newsrooms

A local newsroom has a decisive advantage: it knows in advance which names will keep coming back. They are the ones on its patch, and there are dozens of them, not thousands.

Make the list once. Towns in the distribution area, well-known hamlets, neighbourhoods, rivers. Someone in the newsroom, ideally from the area, reads it through and circles every name whose pronunciation cannot be guessed from its spelling. Usually twenty to fifty remain.

Listen to them in a sentence. A name alone and a name in context are not always read the same way. "The Suresnes town council met yesterday" tells you more than the bare word.

Set them in the account lexicon. An entry set once applies to every later article and takes precedence over any shared base. The procedure is in fixing how a word is pronounced. An account lexicon holds two hundred entries, which is plenty for one territory.

Let the tool flag the rest. Proper nouns in an article are detected before production, and those whose pronunciation looks uncertain are put to you before the audio is made rather than discovered while listening, as described in names flagged before reading. For Breton, Basque and Corsican names, which follow their own logic again, see regional place names in French TTS.

The test to run

Take the local news or "around town" section of your site, where place names are densest. Listen to three articles while following the text, and note every name that makes you wince. If the list goes past five, the voice is not the problem: your lexicon is empty. And if a tool assures you it pronounces every town in France correctly with no setup at all, ask it to read "Sainte-Menehould, Doullens and Saint-Gaudens" in one sentence.

Give your articles a voice with WeDispatch

This blog is itself voiced by WeDispatch. Curious how it sounds on your content?

Book a demo

Read next