Natural French text to speech: the five flaws that still give a machine away

Neural French voices have turned a corner: the metallic timbre, the choppy cadence and the flat intonation of older synthesis have largely gone. Saying a voice "sounds robotic" is no longer true in 2026. But "no longer robotic" is not "indistinguishable from a human". Flaws remain, and knowing them serves two purposes: knowing where to proof an audio before publishing, and judging a tool on evidence rather than on its demo. Here are the five that still give a machine away, said without dressing them up, and, at the end, the ones that have really gone. It extends our work on what it means to read French well.

1. The optional liaison, chosen at random

Compulsory liaison ("les enfants") and forbidden liaison ("les héros") are handled well. It's the optional liaison that gives the game away: "ils ont attendu", "pas encore arrivé". A human speaker makes it or doesn't depending on register, pace and intent, and above all consistently within a single sentence. A synthetic voice decides, but with no logic of register: it may link where a journalist wouldn't, then do the opposite two sentences later. It isn't wrong, it's inconsistent, and that inconsistency is what you hear. This is why liaison remains the hardest test of French read by a machine: there's no single right answer, so no rule can make it perfect.

2. The homograph decided the wrong way

In French, "les poules du couvent couvent" or "il est à l'est" hinge on a word that is spelled one way and said two, with the meaning read from the sentence. Recent voices lean on context and get most everyday cases right, but they still slip on ambiguous phrasing, rare turns and proper nouns that are homographs of a verb. When the model errs, it does so with the same confidence as when it's right, and nothing flags the error on listening: you have to know the text. It's the most insidious flaw, because it doesn't grate, it lies. We cover it in detail in reading homographs.

3. The prosody that flattens on long lists and nested clauses

On a short sentence, the intonation is right. On a sentence with three nested clauses and a list of six items, the melody collapses: every item gets the same contour, and the hierarchy between the essential and the incidental is lost. A human reader slows down, drops the voice on an aside, lifts it on the main clause. The machine reads flat what a badly cut text hands it flat. This flaw is less the voice's than the text's: it eases markedly when the article is written for the ear, with shorter sentences and punctuation that breathes.

4. The reported emotion the voice doesn't act

A neural voice renders a question and an exclamation very well: the pitch rises, the sentence ending is right. It doesn't render irony, reported anger, grief or sarcasm. "Oh, well done", said ironically, will be read as a sincere compliment. A quotation from someone shouting will be read as calmly as the rest. This isn't a missing setting, it's a fundamental limit: the voice reads the literal sense, not the intent. For a news article, which informs rather than performs, that is actually what you want, but you need to know it for texts that live on their tone. We develop the point in what a synthetic voice's emotion conveys, and doesn't.

5. The rare proper noun, read letter by letter

An uncommon place name, a foreign surname, an invented company name: the voice applies French rules to words that don't obey them, and the result is often wrong. This flaw has no perfect automatic fix, because sometimes there's only one right pronunciation and it can't be guessed. The honest answer isn't to pretend the engine knows, it's to correct it once: a pronunciation lexicon fixes the expected form, and the name is then said correctly everywhere. Until someone settles it, the flaw stays.

What, by contrast, can no longer be heard

It would be dishonest to list only the gaps. Several striking flaws have gone. The latent English accent on common words, caused by multilingual models trained mostly on English, is the most spectacular: it was audible on simple words and was fixed by models tuned for French, as we explain in why a French synthetic voice used to sound English. The metallic timbre has given way to a realistic grain. And the flat intonation of simple questions, which rose where it should have fallen, is now right in the vast majority of cases.

The lesson is simple: a 2026 voice is good, not perfect, and progress reached the sound before it reached the sense. That's why our stance remains proofreading before publishing, and targeted correction where a word gives itself away, rather than the promise of a machine that never errs.

Judge a French voice on these five points

The best test is your own, on your own texts. Try it on an article that contains an optional liaison, a homograph and a rare proper noun, and listen to what the voice does with them: ten minutes will tell you what a demo never will.

Read next

• Switching audio provider without losing your audio: the question you forget at signing
• Farm and agricultural press: audio for readers whose hands are busy
• Writing a text-to-speech RFP: what to actually require