Math symbols and operators: what a synthetic voice does with them

A mathematical sign is not a word, yet it still has to be said out loud. « 3 + 2 » reads as "trois plus deux", but « + » on its own, in « les +18 ans » or a « +/- » toggle, does not read the same way, and sometimes should not be read at all. French read by a machine stumbles less on sentences than on these small, densely loaded signs, where writing packs into a single character what speech spreads across three words. It is a cousin of the number problem we cover in reading numbers aloud, but it has its own difficulty: an operator depends entirely on what surrounds it.

The table of signs that trip it up

Here are the forms that show up most often in a general-audience text (a product sheet, an explainer, a recipe post, a comparison), with the reading a human expects and the frequent error of a tool that has not been tuned for French.

| Written form | Expected reading | Frequent error |
|---|---|---|
| 3 + 2 | trois plus deux | "trois croix deux" |
| 5 − 2 | cinq moins deux | "cinq tiret deux" |
| 4 × 3 | quatre multiplié par trois | "quatre x trois" |
| 12 ÷ 4 | douze divisé par quatre | "douze deux points tiret quatre" |
| a = b | a égal b | "a égale égale b" |
| 5 < 10 | cinq inférieur à dix | "cinq chevron dix" |
| x ≤ 3 | x inférieur ou égal à trois | the symbol read as a stray mark |
| ± 2 | plus ou moins deux | "plus deux" |
| √9 | racine carrée de neuf | "v neuf" |
| 50 % | cinquante pour cent | "cinquante pourcent" glued together |
| 3 ‰ | trois pour mille | "trois pour cent" |
| π | pi | the symbol simply skipped |
| 2 × 10⁶ | deux fois dix puissance six | "deux x dix six" |
| ∞ | l'infini | the symbol swallowed without a sound |

Each line is a test you can run against any tool, ours included, in two minutes. The right-hand column is not theoretical: these are the wrong readings you actually hear from generic voices that predict pronunciation character by character, with no model of what a sign means.

Why it is hard for a machine

A neural voice does not compute, it predicts a pronunciation from what it has learned. And the same character changes reading with context. « - » is a minus between two numbers, a hyphen in "quarante-deux", a dash elsewhere. « / » is "divided by" in a calculation, "over" in a fraction, "slash" in a web address. Lowercase « x » is sometimes the letter, sometimes the multiplication sign, and nothing in the character stream says which for certain. A lone operator is trickier still: « + » at the end of "18+" means "and over", not "plus".

Good preparation is therefore not about finding a prettier voice, but about normalising the text before synthesis: recognising that a pattern is a calculation, a comparison, a rate or a notation, and rewriting it into the form the voice will read correctly. That is rule work, and it is exactly the logic of the French text to speech page: what matters happens before the sound comes out. A beautiful voice poorly fed will say "quatre x trois"; an average voice well prepared will say "quatre multiplié par trois".

Where the information was lost upstream

Many failures come not from the engine but from the text handed to it. A power typed « 10^6 » at the keyboard, a « ² » that survived a copy-paste but not the RSS feed that flattened it, a « ≤ » replaced by « <= » on entry: all of these are forms where formatting carried a meaning the character stream then lost. It is the same phenomenon we detail in superscripts, subscripts and scientific notation. The rule is simple: a voice can only read correctly what the text gives it to read.

What WeDispatch does with it, without overpromising

The most common forms in the table (operators between two numbers, the percent sign, per mille, a simple comparison, the square root) are handled by the normalisation applied to every article. That is the baseline, treated as a given rather than an option.

What remains are the cases where context is genuinely ambiguous, even for a human: « + » at the end of a label, an « x » that could be a letter, a house notation specific to a field. There, two honest answers. First, well-entered text removes the ambiguity at the source: writing "and over" rather than « + », "less than or equal to" rather than a sign an export might break, is already half the work, and a good writing habit independent of audio. Second, for a particular reading that must hold across all your articles, the pronunciation lexicon lets you declare the expected form once and for all, with nothing to regenerate. What we do not do: guess the intent behind a notation the text does not make explicit. A machine that decides at random sounds confident, and that is precisely where it gets things wrong with conviction.

Try it on your own texts

The free WeDispatch plan lets you produce the audio version of an article today, and test these cases with your own ears: try it now. The plans, from 69 to 419 EUR excluding VAT per month, are detailed on the pricing page.

Read next

• An audio player in the page or in an iframe: what that choice decides
• Image captions and photo credits: what French text to speech does with them
• Footnotes and note markers: what French text to speech does with them