Aïe, hein, chut: what French text to speech does with interjections

There is a family of words that French text to speech handles badly, and it is not for lack of cleverness: it is because these words follow no rule. Interjections and onomatopoeia (aïe, hein, chut, euh, ouf, hmm, brr, miam) are not spelled like ordinary words. They imitate a sound rather than encoding one, and their spelling is a typographic convention, not a phonetic notation. An engine trained on running text sees few of them, files them wrong, and often renders them in a way that gives the machine away at once.

Why these words are a special case

An ordinary French word is read by applying letter-to-sound correspondences, corrected by context. Château can be decoded, président is settled by grammar. An interjection has no canonical form. People write euh, eu, heu for the same hesitation. They write hmm, hum, mmh for the same thoughtful agreement. The letters are only there to evoke a sound everyone recognises out loud and no one has ever fixed on paper.

So the engine applies its usual rules to a word that was never made for them. Chut comes out as a crisp one-syllable word instead of the expected breath. Hein can surface with a pronounced h, which never happens in spoken French. Grr has no vowel, nothing to hang a syllable on, so the engine stumbles or skips it. These are not isolated bugs: it is the normal behaviour of a system meeting a spelling built for the eye, not for it.

Six interjections and what we expect of them

Here are six common cases, with the reading expected out loud and the error a tool frequently produces.

| Written | Expected reading | Frequent error |
|---------|------------------|----------------|
| aïe | "aïe" in two beats, pain rising | "a-i-e" spelled out, or a dry "aï" with no wince |
| hein | a nasal "in", no "h" at all | the "h" articulated, "h-ein" |
| chut | a breathed "shhh", almost vowel-less | "chute" with a sharp final t |
| euh | a held hesitation, the vowel stretched | a brief "eu", swallowed, losing its hesitating function |
| hmm | a held "m", mouth closed, thoughtful | "aitch-m-m" or three letters spelled out |
| miam | a greedy "miam", one full syllable | "m-i-a-m" segmented, with no lift |

The common thread is obvious: in every case the correct reading depends on an intention (the pain, the hesitation, the silence being asked for) that the spelling does not contain. A machine decoding letters has no way to recover it. This is exactly the kind of finish that separates a voice that reads French from a voice that merely outputs it, and it is the stance we take on the French text to speech page: pronunciation is not decided by letters alone, but by what a human reader knows to do with the text.

What our engine does, and what it does not

Let us be concrete and honest. On the most common interjections, the ones that recur in dialogue and narration, the result is right without any intervention: aïe, oh, ah, bof come through naturally. On rare forms, invented onomatopoeia, and vowel-less consonant clusters (pff, tss, grr), the result is uncertain, and it can vary from one catalogue voice to another. We would rather say so than promise a perfect reading of an object that, by nature, has no single reading.

There is also a question of use. In a news article or a blog post, interjections are rare and usually inside quotations. In fiction, stage dialogue, a transcribed comic, they are everywhere, and that is where the topic becomes genuinely sensitive. If your content is full of them, do not trust a nice demo paragraph: listen to your own hard passages first. It is the same reflex as for pronouncing proper nouns and place names, where only a test on your real text tells the truth.

The move that takes back control

When an interjection matters (a key line, a comic beat that falls flat if the sound is wrong), the guarantee comes not from the voice but from the pronunciation lexicon. You declare there, once, how a specific word should be said, by writing it in a form the engine reads better. Hmm coming out spelled? You fix its reading and the engine stops guessing on that word. It is a targeted correction, not a global setting: it touches only the word you aimed at, across all your content, and it holds over time.

The other lever is the punctuation around the interjection. Chut ! and Chut, dit-elle are not read at the same pace, and that is the wider subject of how punctuation makes speech breathe: a comma, an exclamation mark, an ellipsis change the pause and the intonation, and so the effect. A well-punctuated interjection is already half well read.

How to test a tool in two minutes

Take five interjections your content actually uses and have them read in a row: Aïe, Hein ?, Chut, Euh, je ne sais pas, Hmm, peut-être. Listen for three things: does the h in hein stay silent, does chut breathe instead of snapping, is the hesitation euh held. A tool that gets all three handles spoken French, not just written French. A tool that spells out hmm tells you everything you need to know before committing to an audiobook or a work of fiction.

Interjections are not a purist's detail. They are the words where the ear detects fastest that audio was laid on carelessly, because they are the words we recognise without thinking. One botched interjection in a line of dialogue, and the listener feels the machine behind the voice.

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