Does text to speech learn from your corrections?

You listen to your article before publishing, you hear a town name mangled, you fix the passage. The next question comes straight after: does that fix count for anything, or will you redo it tomorrow on the next article that contains the same word? It is a fair question, and the answer is a fair one too. Yes, a reading correction is kept, but not in the slightly magical way people imagine. It does not vanish into a model that quietly improves on its own. It follows a precise path, in three stages, and only one of them is automatic.

What a reading correction is, and what it is not

Not everything can serve as a signal. If you change two words in a sentence, you have rewritten your text, not corrected a pronunciation, and the product should draw no conclusion from it about how French is read. So WeDispatch only keeps a short, clean correction: a single word changed in a brief passage, with a hard limit of four words and about sixty characters. Beyond that, it is a rewrite, and it is discarded.

A second, subtler filter protects other newsrooms from a blunder. A reading correction almost always keeps the word's first letter, or lengthens it to spell it out: "Gyökeres" becomes "Gueukeurech". If you replace "Rennes" with "Nantes", the product refuses to treat it as a pronunciation fix, because it is not one: it is an editorial change of substance, and it would be serious to offer it to someone else. That caution comes from a stance we hold throughout the way we handle French text to speech: when in doubt, leave it alone.

Stage one: for you, right away

A one-word correction becomes a proposed lexicon entry, for your account alone. You accept it with one click, and from then on the word is said correctly across all your following articles, without your having to think about it again. It is the natural extension of the pronunciation lexicon you can also fill in by hand.

Nothing is written without your click, and that is deliberate. A lexicon entry is an editorial call: deciding that your town is said this way and not another belongs to you, not to the machine. The product suggests, it does not decide for you. The most frequent corrections rise to the top of the suggestion list, so the one gesture of accepting goes where it is most useful.

Stage two: between newsrooms, under a condition

Here is the part that surprises people, and the only one where your correction can serve others. If three independent newsrooms correct the same word the same way, it is no longer a local preference, it is probably a rule of French we did not yet know. The word then becomes a candidate for the shared dictionary.

Two guardrails frame that step, and they matter. First, the count is on the number of distinct newsrooms, never on the number of corrections: a single newsroom that fixes the same name thirty times is expressing its preference, not a general truth. Second, a candidate never enters the dictionary on its own. A human on our side has to review and validate it, because one wrong shared rule would poison the whole estate at once, with no one able to tell where the fault came from. What the product builds at this stage is a list to be reviewed, not an active rule.

And what travels between accounts is deliberately tiny: the pair "mispronounced word to correct word", never the sentence, never the article. Your text does not leave your account. The signal always comes from the newsrooms that listen, never from the readers: no listener data enters this mechanism at all.

Stage three: what is learned does not break again

Every confirmed correction becomes a case in a test bench we keep running. The consequence is undramatic and the most important of the three: the product can no longer reintroduce an error it has corrected once. A word said well today stays said well after the next engine update.

What it does not do, and why that is on purpose

This is not a model that learns, it is a count over human corrections. That choice has a cost and a benefit. The cost: the product will never invent the pronunciation of a word it has never seen corrected, it waits for your gesture. The benefit: everything stays legible, reproducible and explainable to a committee, which is no small thing when you have to account for your production chain against an internal policy. It is the same logic as our refusal of an engine that would read French without a verifiable rule. A correction that truly serves is worth more than a promise of intelligence that improves without anyone being able to say how.

Read next

• Offer your articles as audio in another language, without redoing the work
• An English quote in the middle of a French article: what the voice does with it
• Add an audio version to a Google Sites site