French text-to-speech for children: choosing a voice for young listeners

Some of the most thoughtful requests we get come from people writing for young readers: a French-language children's news magazine, a public library publishing reading picks, a school sending notes home to families, a science outreach group working with 8 to 12 year olds. They all ask a version of the same question: can a French synthetic voice work for children, and if so, which one?

Yes, within limits that are worth knowing before you start. Here is how we think about it, including the parts where our answer is "use a human instead".

We do not offer a child's voice, on purpose

The first request is often for a child's voice, "so kids can relate". We do not offer one, and the reason can be checked in our voice registry.

That registry lists every voice our speech engine exposes, with its language, the vendor's description, and the date a person on our side listened to it. Four voices in it are described as child voices or picture-book narrators. They are English and Japanese. None is French, and none is enabled. Every French voice we offer is an adult voice.

The rule behind the registry does the rest: no voice reaches a customer until someone has listened to it and signed off, with a date. We have not heard a French child voice good enough to read a full article without becoming tiring or gimmicky, so we offer none rather than a novelty that sounds "childlike" for thirty seconds.

There is also a simpler point. Children mostly do not listen to children. They listen to a parent, a teacher, a storyteller, a radio presenter. In children's audiobooks, the narrator is almost always an adult, and child voices appear mainly to play characters, which is precisely what an article-reading tool does not do.

Which adult voice to pick

Each French voice in our catalogue carries a short style note. For young audiences, two groups behave very differently:

• warm, even voices (Pauline, noted as "warm, close", or Michel, "rich, steady") hold up over several minutes and feel reassuring;
• confident, corporate or phone-style voices, excellent for a market update, quickly sound like voicemail to a seven-year-old.

Male versus female matters less than people expect. What tires a young listener is a clipped delivery and flat intonation. Our wider criteria are in choosing a French voice by content. In practice, play the same page with two or three voices and let your ears decide.

Speed belongs to the listener

A child does not follow text at the pace of a busy adult. The player offers a 0.8x speed below normal, and the listener's choice is remembered from one article to the next without an account. See how playback speed works.

That setting helps, but it does not fix writing. A dense paragraph played slowly is still dense, just longer.

What a reading voice cannot act out

This is the limitation people most often discover too late.

Characters. A storyteller gives the wolf, the pig and the narrator different voices. Our tool reads text; it does not perform it. It can switch voice for quotations in an article, but it will not cast five characters in a fairy tale. How quotations are detected is explained in multiple voices for quotes.

Acted emotion. Questions and exclamations come through well. Rising fear, a deliberately silly joke, a whisper: much less so. We go into detail in what synthetic voices convey of emotion.

Sound effects. French "badaboum", "miaou" or "pschitt" are read as words, not as noises. A story that leans on them loses a lot when heard.

So the practical line is clear. Synthetic speech suits non-fiction for children very well: a news story explained for kids, a page about an animal, a write-up of a class trip. It suits a storybook that lives through its characters much less well. For that, a human reading is still the right call, and we are happy to say so.

Writing for an eight-year-old's ear

The usual rules for audio-friendly writing apply, only tighter:

• one idea per sentence, no asides in the middle;
• few numbers, rounded: "almost two hundred" rather than "197";
• acronyms spelled out the first time, and again later in a long piece;
• tricky proper names (a foreign scientist, a faraway city) checked by ear before publishing and fixed once in the pronunciation lexicon.

More in writing for the ear. For adults these are good habits; for children they are the difference between listening and drifting off.

The words on screen matter as much as the voice

Early readers listen and look at the same time. Guided reading, which highlights each word on screen as it is spoken, serves them directly, as it does readers with dyslexia. It depends on word-level timing that not every speech engine provides; what it takes is laid out in word-by-word highlighted reading.

For a school or library site, the next question is listener data for minors. What our player records, and what it does not, is covered in audio for school websites, and libraries will find more in our piece on reading accessibility in media centres.

A test worth ten meetings

Take a real page from your site, ideally one with a proper name, a number and a line of dialogue. Have it read by two voices, play it at normal speed and at 0.8x, then play it to a child of the right age and watch. Their reaction will tell you more than any vendor pitch, ours included. What we stand for when it comes to reading French well, and what we deliberately do not do, is set out on our French text-to-speech page.

Read next

• 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
• When your article audio service goes down: what readers actually see