5G, COVID-19, A320, CAC 40: how French text-to-speech handles words built from letters and digits

Take any French news article and look for words that combine letters and digits. They are everywhere: la 5G, le COVID-19, l'A320, le CAC 40, le G7, la RN 7, TF1, la PS5, la COP 28. Each has an established spoken form, and no single rule covers them. "5G" is said "cinq G", digit first. "A320" is "A trois cent vingt", letter first. "COVID-19" is not spelled out; it is pronounced as a word and followed by the number. "TF1" is spelled, then "un".

For a synthetic voice these are a particular kind of trap. A bare acronym can be spelled or pronounced, and we have covered elsewhere how French acronyms and abbreviations get decided. A bare number is easy. But once the two are fused, the reader has to work out where one ends and the other begins, which order to say them in, and whether the digits are a quantity ("forty") or a label.

What we checked, and how

Rather than describe what we intend, we ran around thirty news sentences through our text preparation, the same step every article goes through before it is read. It does not make the voice speak. It rewrites, when it knows how, what the voice will receive. The results fall into three groups.

What the preparation rewrites

| Written | Sent to the voice | Rule that applies |
|---|---|---|
| le CAC 40 | le CAC quarante | a detached number is said in words |
| la RN 7, la RN-7 | la erre enne sept | a road, recognised from context |
| l'A7 est fermée | l'a sept est fermée | the elided article marks a road |
| la D 939 | la dé neuf cent trente-neuf | same, for a departmental road |
| GPT-4 | gé pé té quatre | spelled acronym fused to its number |
| la TGV-2.1 | la té gé vé deux point un | dotted version number |
| un 4x4 | un quatre-quatre | usage, not multiplication |
| le CO2, le H2O | le cé o deux, le ache deux o | a closed list of chemical formulas |
| la COP 28 | la COP vingt-huit | detached number again |

What these share is a clue. For roads, the rule requires a road word in front ("autoroute", "nationale") or the elided article ("l'A7", "sur la N7"). A lone "A7" could be a product code or a spreadsheet cell, so the rule stays out of it. For chemistry, the list is closed: CO2, H2O, NO2, SO2, O2, O3, CH4, N2O. Anywhere else, a letter stuck to a digit is a reference, and we leave it alone.

What the preparation leaves to the voice

5G, COVID-19, "un A320", G7, F1, TF1, PS5, COP28 written without a space, MP3, the "loi 3DS": nothing is rewritten. That is a choice. For "COVID-19" the code states it plainly: an acronym pronounced as a word stays with the engine, which reads it far better than any spelling we could substitute. For the rest, no clue in the sentence separates the possible readings with enough certainty. Rewriting blind would turn the voice's hesitation into a mistake with our name on it.

One small case shows why we stay cautious. "Sur M6 ce soir" (on M6 tonight, the TV channel) comes out as "sur emme six ce soir". The result is right, but it was produced by the road rule, because M is also a letter used for metropolitan roads and "sur" precedes it. "Il regarde M6 ce soir", with no such clue, goes through untouched. The aircraft gets the same treatment: "l'A320" is taken for a road because of the elided article and comes out "l'a trois cent vingt", which happens to be right, while "un A320", with no elision, goes through as written. Both times the right reading rests on a coincidence, and a coincidence is not a rule.

The one that still fails

One spelling still comes out badly: "la RD-939", hyphenated with a three-digit number, gives "la erre dé-939". The acronym is spelled but the number stays glued to the hyphen, because the fused-number rule only takes two digits and the road rule does not accept a hyphen. Written "RD 939" with a space, it reads correctly. We mention it because a table like the one above is only believable with its misses in plain sight: a preparation step with no failures has not been tested.

How your newsroom settles the rest

For everything left to the voice, the call is yours, and you make it once. The pronunciation lexicon on your account accepts words that contain digits, written as spelling = pronunciation:

• A320 = A trois cent vingt
• 5G = cinq G
• PS5 = pé esse cinq

We checked those three entries: they apply like any other word, whole word only, and they run before the automatic rules, so your decision always wins over ours. The displayed text, subtitles and read-along keep the article's spelling. "A320" stays "A320" on screen; only the voice receives the pronunciation. If you are new to the lexicon, our guide on fixing how a word is pronounced covers the method.

The sensible habit is not to put everything in the lexicon. Listen first: a good French neural voice reads many of these tokens correctly on its own, and a pointless entry is one fewer slot for the cases that matter. Only fix what you have actually heard go wrong, on words that keep coming back: aircraft models for an aviation title, bus and tram lines for a regional daily, product references for an online shop.

Five sentences for a quick check

If you are evaluating a tool, or want to know where yours stands, this set takes a minute to hear and covers all three groups:

• "Le CAC 40 a perdu 1,2 % à l'ouverture."
• "Bouchons sur l'A7 et sur la RD 939."
• "La 5G arrive dans le village."
• "Un A320 de la compagnie a été cloué au sol."
• "Le COVID-19 circule encore cet hiver."

The first two test what a preparation step can know; the last three test what the voice does unaided. That split matters more than any overall score, because it tells you what you will have to correct yourself. It is also the line we hold on our French text-to-speech page: rewrite when a rule allows it, hold back when it does not, and give you control over the rest.

Read next

• Tracking audio listens in Google Analytics 4 or Matomo: what the player sends, and how to catch it
• Quote, unquote: should a text-to-speech article announce its quotations?
• Changing your audio volume mid-subscription: what happens to your credits and your invoice