French election results read by text to speech: a test on percentages, vote counts and party acronyms

If you publish in French, or run a bilingual desk that does, the next big test for your audio versions has a date: France's presidential election in spring 2027, and the local results pages that come with every vote. Election copy is short, formulaic and published fast. It is also where a synthetic voice is most exposed, because nearly every sentence carries a number, a sign or an acronym, and French reads all three differently from English.

We did not want to answer "will it read our results correctly?" from memory. So we ran twenty-one typical sentences through the text preparation stage that sits in front of the voice, the stage that decides how "38,42 %" or "RN" will be said. What follows is the actual output, copied without edits.

What French does differently

Three conventions trip up tools built mostly for English, and an English-speaking editor checking the audio may not notice them:

• The decimal mark is a comma. "38,42 %" is thirty-eight point four two percent, read as "trente-huit virgule quarante-deux pour cent". A tool that treats the comma as a list separator reads two numbers.
• Thousands are separated by a space. "12 345 voix" is one number, twelve thousand three hundred forty-five votes. Read naively, it becomes "twelve, three hundred forty-five".
• Acronyms split into spelled and spoken. French says "le PS" letter by letter, but "le Modem" as a word. There is no reliable rule from the letters alone.

Percentages, counts and rounds: what works

This is the heart of any results story, and it is the solid part.

• "Au 1er tour, la liste conduite par Mme Durand obtient 38,42 % des suffrages exprimés" becomes "Au premier tour, la liste conduite par Madame Durand obtient trente-huit virgule quarante-deux pour cent des suffrages exprimés". The ordinal, the title and the decimal are all handled.
• "Participation : 61,7 % à 17 h, contre 58,3 % en 2020" comes out with the time read as "dix-sept heures" and the year as "deux mille vingt".
• "12 345 voix" becomes "douze mille trois cent quarante-cinq voix": the space is understood as a thousands separator.
• "2nd tour" is read "second tour" and "2e tour" "deuxième tour", keeping the distinction the journalist wrote.
• Polling margins work too: "27 % (+2), marge d'erreur de ± 3 points" is read with "plus deux" and "plus ou moins trois points".

The general rules for rates are in our piece on how French TTS reads percentages.

Party acronyms

"RN" comes out as "erre enne", "PS" as "pé esse", "LR" as "elle erre", "NFP" as "enne effe pé", "DVD" (the French label for miscellaneous right-wing candidates, not the disc) as "dé vé dé", and "EELV" as "eu eu elle vé". "Modem" is left alone and pronounced as a word, which is correct.

Two honest caveats. First, writing letters out phonetically has a known weakness on two-letter acronyms: our own development notes record that "UE" rewritten as "u eu" was read as a word, so it is now left untouched for the voice to spell. PS, RN and LR are short in the same way and are worth a listen. Second, "LFI" and "UDI" are passed to the voice unchanged, after a rule that lowercases surnames written in capitals turned "le député LFI" into "Lfi". We do not claim to know how every voice in the catalogue handles them. How the tool chooses between spelling and pronouncing, and how to override it, is covered in our article on acronyms in French TTS.

The two errors

A detached plus sign. "(+ 812 par rapport à 2020)", with a space between the sign and the number, comes out as "(+ huit cent douze …)". The sign is not turned into a word, so what you hear depends on the voice. Fix it in the copy: attach the sign, or write "812 de plus qu'en 2020".

A hyphen between two results. "52,1 % - 47,9 %" comes out as "cinquante-deux virgule un à quarante-sept virgule neuf pour cent". The preparation recognised two numbers joined by a dash and read it as a range or a sports score ("52-48" is rightly "cinquante-deux à quarante-huit"). For a runoff, "à" is wrong: a French listener expects "contre", and the first figure loses its "pour cent". Written with a word, "50,8 % contre 49,2 %", the sentence is read exactly as intended. One style rule for the night: never a dash between two scores.

What to prepare before the night

Results go live within the hour, and nobody will re-listen to each one. What works is preparation the week before:

• Candidate names. They are the words a voice knows least, and the most visible when wrong. Set the pronunciation once in the pronunciation lexicon and every later article uses it.
• The results template. If the desk uses a standard brief for each town, listen to the template. The phrasing repeats, and so do the errors.
• The two style rules above: "contre" between scores, and signs attached to their numbers.

For how the whole chain works between your text and the voice, and what we deliberately do not do, see our page on French text to speech.

The scorecard

Across twenty-one ordinary sentences: percentages, vote counts, ordinals, times and margins of error are read correctly. Party acronyms are spelled and Modem is pronounced. Two real errors, both fixable by a writing habit rather than a setting. Two acronyms, LFI and UDI, are left to the voice and should be checked by ear before election night rather than during it.

Read next

• Domain verification for article audio: why a text-to-speech service should ask, and how ours works
• 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