Superscripts, subscripts and scientific notation: what a synthetic voice does with them

A feature article, a spec sheet, a health explainer or a research post almost always contains forms that are not plain words: an area in square metres, a chemical formula, a power of ten, an ordinal, a temperature. On the page, you grasp them at a glance. Read aloud, they force the engine to decide: is this a number to say, letters to spell, or a symbol to skip? That decision is where many texts quietly betray their author, because the reading you expect and the reading you get do not match.

A string of characters has lost what the formatting was saying

When you type "m²", your eye sees a unit: the square metre. To a reading engine, "m²" is a sequence of two characters, a letter and a Unicode superscript symbol. Nothing in that sequence says the "2" is a power rather than a footnote marker or an issue number. The formatting held that information: it placed the 2 small and high. The text handed to the synthesis has often lost it along the way, especially if it went through a copy paste, an export or an RSS feed that flattens everything.

So this is not a question of voice quality. It is a question of available information. A voice can only read correctly what the text gives it to read.

What passes, and what trips up

Some forms are so common they are learned and read effortlessly. Units written out in full, or very widespread abbreviations like "km" or "kg", rarely cause trouble. The difficulties begin with superscripts, subscripts and condensed notation.

Here are the cases that come up most often, with the reading a human listener expects and the frequent error of a tool that has not been tuned for it:

• 120 m²: expected "one hundred and twenty square metres"; frequent error "one hundred and twenty m two".
• H2O: expected "H two O" (or "water" depending on context); frequent error "h twenty" when the 2 is glued on.
• CO2: expected "C O two"; frequent error "co two" read as a single word.
• 10^6 or 10⁶: expected "ten to the sixth" or "one million"; frequent error "ten six" or "ten caret six".
• 20 °C: expected "twenty degrees Celsius"; frequent error "twenty degree C" or dropping the symbol entirely.
• x²: expected "x squared"; frequent error "x two".
• 1er, 2e, XIXe (in French text): expected "first", "second", "nineteenth"; frequent error "one e r", "two e", "X I X e".

The common thread is simple: a superscript or subscript digit has two possible readings, and with no contextual clue the tool takes the most literal one. That is cautious, but to the ear it is a mistake.

Why the "2" is a special trap

The square and the cube dominate everyday text, yet they are written with the same digits as any other quantity. When you read "an area of 50 m²", your brain knows the 2 belongs to the unit. A machine sees a 50, a space, an m and a 2. If that 2 is a genuine Unicode superscript character (²), good preprocessing can recognise it; if it was flattened into an ordinary 2 by a conversion, the information is gone for good and no reading can reliably guess it back. The same quantity written "50 m2" and "50 m²" therefore does not read the same, even though it looks the same.

This is exactly the logic we lay out for numbers in general in our piece on reading numbers aloud, and for ordinals in the one on Roman numerals: the written form decides the reading, not the intent.

Taking control when the default reading is not enough

Our stance is to read French well rather than read thirty languages roughly, and that stance is decided precisely on these edge cases. Two levers exist, and they complement each other.

The first is editorial and free: write the readable form when accuracy matters. "square metre" rather than "m²" in a text meant mainly for listening, "carbon dioxide" rather than "CO2" in a general article. It is not always desirable (a spec sheet loses density), but it is the most reliable setting, because it depends on no tool.

The second is the pronunciation lexicon: you declare once how a recurring form should be read, and the tool applies it everywhere. A scientific journal that writes "CO2" a hundred times a month does not need to rewrite its text: it sets the reading once. The exact move, and its limits, are described in our guide on targeted pronunciation fixes. It is also the right place to check what French text to speech does by default before adding a rule.

In practice, for a newsroom

Three habits are enough to avoid almost every unpleasant surprise. First, prefer real superscript and subscript characters to a plain glued digit when your editor allows it, because they give the tool a chance to recognise the form. Then, always listen to the passages that contain formulas, units or ordinals: those are what break, never the ordinary sentences. Finally, for forms you use often, declare them once in the lexicon rather than fixing them by hand article after article.

The exponent and the subscript are not a flaw in the voice: they are places where the text has to say what it means. Once it does, the reading follows.

Listen to your own formulas

Take one of your most technical articles, the one full of units, formulas or percentages, and run the test. Within two minutes you will know which of your writing habits read cleanly, and which deserve a reading rule.

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