French text to speech for learners: a useful listening partner, not a teacher

Ask anyone who learned French as an adult what tripped them up, and spelling rarely comes first. The problem is the distance between the page and the ear. Half the letters are silent, words run into each other, and "les enfants" picks up a "z" sound that is nowhere on the page while "les haricots" pointedly does not. A learner can read a paragraph perfectly well and fail to recognise the same paragraph spoken aloud.

Text to speech can narrow that gap, because it makes any written French audible, on demand, as many times as you like. This post looks at what it actually teaches, where it misleads, and how a teacher or a language school can judge a tool before pointing students at it. Our own product is used as the worked example, limits included.

Where it helps: tying letters to sounds

The most valuable exercise is the simplest one: read and listen at the same time. When an article is narrated with word-level alignment, the word being spoken lights up on screen as the voice reaches it. We call this guided reading, and for a learner it removes the main friction of listening with a text: losing your place.

The second benefit is easy to overlook: consistency. A synthetic voice says the same sentence the same way every time. Native listeners find that tiring over a long piece. A learner finds it reassuring: they can loop one sentence ten times and the intonation never shifts, which no human recording can promise from one take to the next.

The third is coverage. No teacher can record every handout, and no news site will hire an actor for every article. Synthesis makes audible what would otherwise stay silent.

Liaison: a solid model where it matters

Liaison is the classic learner trap, and also the hardest test for a machine, which is why we built a short test around it in French liaison, the hardest test for TTS. For learning purposes, the results break down like this:

• Compulsory liaisons ("ils en ont assez", "les anciens élus") are made every time, so the learner hears the correct form.
• After "et" there is never a liaison, and this is handled as a rule rather than left to chance.
• Before an aspirated h ("les héros", "les haricots") common words are handled; for a rare word, the guarantee comes from a human entry in the pronunciation lexicon, not from the voice.
• Optional liaisons ("pas encore") are read in one register and kept consistent through the whole text.

That last point needs a teacher's note. The voice does not signal that a liaison is optional; it simply picks one. A student who only ever hears that voice may conclude there is a single correct way to say "pas encore". There are two, and someone has to say so.

Slowing down is not the same as articulating

Every learner wants a slower voice. The player allows it: the reading-comfort panel goes down to 0.75x, and the site owner can also set a slower default speed (0.9x). The listener's choice is remembered between articles without any account.

What this does, precisely, is play the same file more slowly. It does not insert pauses between words or separate syllables. Linked sounds stay linked, just stretched. That is arguably a good thing: the learner hears real connected French rather than a dictation voice nobody actually uses. But a complete beginner hoping for "word, pause, word" will not get it. The distinction between listener speed and the voice's own pace is covered in audio player playback speed.

What it does not teach

This is where tool marketing tends to overreach, so plainly:

• It does not correct the learner. A text-to-speech voice talks; it does not listen. Assessing a student's pronunciation is speech recognition, a different technology.
• It offers one norm. Our French voices are designed for the French of France. None is labelled with a regional accent, and none sounds like Montreal, Brussels or Dakar. We went into this for Quebec French. For a learner living in Quebec, the model is useful but incomplete.
• It does not perform emotion. Irony, irritation, warmth carried by tone rather than words come out neutral, and that is exactly what advanced learners struggle to catch.
• It explains nothing. It reads "les haricots" without a liaison; it does not tell you why.

Who gets real value from it

For a language school or a newcomer-support organisation, the most direct use is narrating what you already publish: news on your site, practical guides, instructions. Our French text to speech page describes how the text is prepared before it is read, and anything that trips the voice (a place name, a technical term) is fixed once in the lexicon and stays fixed.

For sites written in plain French for new arrivals, audio adds to the writing work rather than replacing it. A short text with simple sentences sounds far better than a dense one played slowly. The overlap with easy-to-read writing is discussed in our post on cognitive accessibility.

A five-sentence test before you recommend a tool

Take five sentences from your own course material: one with a compulsory liaison, one with "et" followed by a vowel, one with an aspirated h, one question, and one number such as "1 280 euros". Have the tool read them while you follow the text, then slow it down. If the first three are right, the question rises at the end and the number is said the way you teach it, the voice is a good listening companion. It will never be a teacher, and nobody should sell it as one.

Read next

• Article audio in your own player: keeping your design system and your security policy
• Audio-ready webhooks: stop polling and let the API tell you
• French text-to-speech on product pages: what gets read right, and what doesn't