Flagging uncertain names before the reading, not after

The moment you hear a name read wrong always comes too late. The audio is made, it has been paid for, and fixing it means running the reading again. Most tools let you find the mistake by ear, once the file is ready. WeDispatch takes the problem from the other end: it flags the names whose pronunciation is doubtful before it reads, lets you settle each one once, and never asks you the same question twice.

How a word becomes a candidate

Spotting the proper nouns in an article without a dictionary rests on a simple fact about written French. A word like "Le", "Les" or "Cette" is only capitalised at the start of a sentence, after a full stop. A proper noun keeps its capital in the middle of a sentence. So a capitalised word that appears anywhere other than at the start of a sentence is, almost certainly, a proper noun. That single rule sets aside the vast majority of ordinary words, without having to know any of them.

The rest is caution. Runs of capitals are grouped ("Diên Biên Phu" counts as one entry, not three, since three separate entries would each get a pronunciation decided in a vacuum). Months, weekdays and courtesy titles, capitalised out of habit, are excluded outright. And the number of candidates per article is capped: beyond a point, you are no longer asking for an opinion, you are demanding an inventory.

What applies at once, and what waits for you

For the hard names, the tool proposes a pronunciation. It sorts them into two levels, and that dividing line is a product decision, not a detail. A high-confidence proposal applies at once to your account's lexicon, and it undoes in a single click on your screen: it is there, visible, reversible. A low-confidence proposal does not apply; it is presented plainly as a doubt, held in a queue, and nothing enters your lexicon until you have approved it. When in doubt, we doubt: anything not explicitly certain is treated as uncertain.

You are never asked to judge a pronunciation from text. A phonetic spelling means nothing on a screen. Before you decide, you hear the word as the voice says it today, then as it would be corrected, and you choose by ear. What is at stake here goes to the heart of our approach to French text to speech: a pronunciation is not read, it is heard.

Where it shows up

Proposals appear in two places. On the article's "Pronunciation" card, sorted most-frequent first, because a name repeated ten times costs more to leave unaddressed than one mentioned once. And in the "your call" column of My audios, which tells you at a glance which articles are still waiting for a decision. What you settle there is invisible on the reader's side: your text stays intact on screen, the read-along and the subtitles show the article as you wrote it, never our pronunciation spelling. Only the engine receives the correction, and it applies before every synthesis, including when you are in charter mode: it is a rule, not a black box.

And if nobody decides

Nothing breaks. An article is voiced perfectly well without a single proposal being approved. A low-confidence item left waiting changes nothing about the reading, which keeps the default pronunciation; a high-confidence one is already applied, and stays reversible until you confirm it. The flag is a convenience, and a convenience that broke the making of an audio would no longer be one. So you can publish without deciding anything, and come back to settle a name later, once it has bothered you on listening.

Why a name everyone says the same never comes back

Two names do not have the same status. "Zelensky" is pronounced the same way in every newsroom: it is a universal name, its pronunciation is decided once and for all, in a base shared across every account, and it is not put back to you afterwards. "Le Guern", the surname of a local councillor, may be said differently from one region to the next: it is editorial, it stays your account's business, and nobody decides in your place. When the category is unclear, a name is treated as editorial, because a name wrongly filed as universal would change the reading for everyone.

There is one last, quieter filter: a proposal identical to the word teaches nothing. If the voice already says a name correctly, no lexicon entry is created, and the name is not flagged to you. So you only ever see what deserves a decision, never the background noise of words that come through fine on their own. It is the natural companion of the pronunciation lexicon, which lets you correct a word by hand: flagging upstream tells you which ones are worth the trouble, and the work on proper nouns and place names becomes a short queue rather than a blind hunt.

Read next

• One podcast per section, each configurable, on Apple Podcasts and Spotify
• Reading a PDF aloud: what grates on the ear, and what we do with it
• Listen to your articles and PDFs on your phone, like a radio