Asking a question out loud: what French text to speech does with intonation

In French, a question is not always recognised by its grammar. It is recognised by its melody: the voice rises at the end. This is true out loud long before it is true on the page, and it is exactly what makes the task hard for a machine. Text to speech has to decide, sentence by sentence, whether to end falling (a statement) or rising (a question). When it gets this wrong, the ear catches it at once, before you even work out why: a question asked on a flat tone rings false, like a badly delivered line.

Three ways to ask the same question

French inherited three interrogative forms, and they do not coexist by accident: they carry neither the same register nor the same intonation load.

| Written form | What flags the question | Expected intonation |
|--------------|-------------------------|---------------------|
| Viens-tu ? | subject and verb inverted | a clear rise on the final syllable |
| Est-ce que tu viens ? | the phrase "est-ce que" up front | a moderate rise, tension set from the start |
| Tu viens ? | nothing, except the question mark | a sharp rise, and it alone makes the question |

The third row is the heart of the problem. "Tu viens" and "Tu viens ?" are the same string of words. Remove the question mark and you have a perfectly correct statement. So it is the punctuation, and it alone, that must trigger the rise. A tool that reads the words without looking at the punctuation will read this question as a flat line, and the most common spoken form (the intonation-only question) will be precisely the worst rendered.

This is one of the places where our stance shows, the one we defend on the French text to speech page: reading a language well is not decoding letters, it is recovering what a human reader does without thinking. Nobody needs to be told to raise their voice on "Tu viens ?". The machine does need to be taught.

What our engine does with a question

Let us be concrete. The question mark is the primary signal, and it is taken into account: across the three forms in the table, the end of the sentence rises because the "?" is there. The intonation-only form, the most fragile in theory, is therefore rendered correctly in practice, on one condition: that the question mark is written. That is the single most useful reminder in this whole article. A question without its mark ("Tu viens", ending on a full stop or nothing) will be read as a statement, because nothing tells the voice to rise.

Where the result stays imperfect is on what writing does not contain. A long question, with several clauses before the verb, needs a curve that builds and holds back, and the engine tends to play everything on the last syllable alone. A rhetorical question ("Faut-il vraiment le rappeler ?") calls for a particular intonation, often less rising, because it expects no answer: that nuance an engine cannot guess, lacking any grasp of intent. We would rather say so than let you believe in a reading that seizes meaning. What is seized is the punctuation; the rest is an honest approximation.

The move that takes back control

Punctuation is your first lever, and the simplest. Write the question mark, always, even on a casual question. Split an over-long question in two if its curve becomes unreadable: two well-formed questions beat one the voice cannot sustain. It is the direct extension of how punctuation makes speech breathe, where each mark carries a different pause and intonation.

The second lever is about dialogue, where questions come one after another. "Tu viens ?, he asked" and "Tu viens ? He was waiting" are not read at the same pace, and the clause that follows changes how the question settles. That is the wider subject of dialogue and quotations: the voice does not treat an isolated question the way it treats one caught in an exchange.

How to test a tool in two minutes

Take three sentences and have them read in a row: "Viens-tu ?", "Est-ce que tu viens ?", "Tu viens ?". Listen for one thing only: does the voice rise at the end of all three. Then add the same string of words without a question mark ("Tu viens.") and check that it falls: that is the proof the tool really reads the punctuation, rather than slapping a rise on at random. A tool that rises on the three questions and falls on the statement handles French interrogation. A tool that reads "Tu viens ?" like "Tu viens." tells you it is decoding words without listening to the sentence.

Question intonation is no stylist's flourish. It is what separates a voice that reads a text from a voice that understands it enough to ask the question on your behalf. And as so often with reading French, most of the work falls to you first: a well-punctuated question is already half well read.

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