← Blog · 18 November 2026 · Lire en français

Text to speech: the voice says one thing, the transcript must show another

A good French voice says « mille soixante-six » where the article prints 1066. The transcript, the read-along text and the subtitles should still show the article. How we keep two versions of the same text, what broke in September, and what happens when alignment fails.

Every text-to-speech tool that reads a language properly rewrites it first. In French, "1066" has to become mille soixante-six, "SNCF" has to be spelled letter by letter, a football score like "23-11" needs a spoken connector, and a foreign surname fixed in the pronunciation lexicon becomes a phonetic spelling nobody would ever print. We have written separately about numbers and acronyms. This piece is about what happens to that rewritten text afterwards.

Because an article's audio is not only sound. There is the transcript under the player, the read-along text highlighted word by word while you listen, and the SRT and VTT subtitle files. All of them are built from the word timestamps the engine returns with the audio. And those timestamps describe the words the engine said, not the words the journalist wrote.

A defect you cannot hear

Until 5 August 2026, the text we sent to the engine and the text we displayed were one and the same. Nobody noticed by listening, since the sound was right. But the read-along view showed the phonetic respelling of a Swedish footballer's name instead of the name itself, a score written out in words, and an acronym rendered as four spelled letters.

In other words, readers were not looking at the article. They were looking at our pronunciation draft. To them it read like a spelling mistake by the newspaper, on a word the voice pronounced perfectly. This kind of problem never comes back through listening checks; you only catch it with your eyes on the screen while the audio plays.

One pipeline, two faces

The fix rests on one rule: every transformation in our text preparation keeps both of its faces. When a rule turns "1066" into the spoken form, it does not overwrite the digits. It wraps the passage in a pair of markers that hold both versions, the spoken one and the displayed one. At the end, the same text is resolved twice, once for the engine and once for the screen.

After synthesis, a second step realigns the timestamps. The three spoken words are merged into a single displayed word, "1066", which takes the start time of the first and the end time of the last. Highlighting stays accurate to the millisecond and the reader sees the author's text. The read-along view and the subtitle files inherit this at no extra cost.

Two design choices matter more than they seem:

  • Markers, not opaque tokens. An earlier version replaced each passage with an opaque token, which made it invisible to every later rule. Measured on 19 September, 8 of the 71 cases in our reading test bench came out differently in production than on the bench; "30° in the shade" ended up half converted. Markers leave the passage readable, so later rules can still work on it.
  • A rule that would split a marker pair is refused. The text keeps its original form. Losing one transformation is better than ending up with two faces that no longer match.

When it broke, with the numbers

The idea held; the realignment did not, at first. It searched for each passage within forty words of the previous one, and when a passage sat further away, it gave up without moving forward, so every following passage was dropped too. Measured in production on 28 September: 21 audios out of 21 affected, and in one long column, 344 passages out of 348 left in their spoken spelling. The sound was right everywhere. The screen was not.

The same-day fix gives every passage its expected position, compares letters rather than punctuation, picks the candidate closest to where it should be, and keeps French elisions intact on screen. On a dense news article: 200 passages realigned out of 204 with a perfect engine, zero misplacements. On a column where the engine drops 5% of the words: 39 out of 39, against 2 before.

For audio already produced, regenerating would have charged customers credits for a display bug. So we wrote a repair that replays the realignment on the stored words: a new version with the same sound, copied as is, and the transcript returned to the author's spelling. No synthesis, no credits, and the previous version stays in the history.

When realignment fails

Engines do not always split words the way we do, and they occasionally swallow one. When that happens we do not guess. The timestamps stay as the engine returned them, that passage shows its spoken spelling, and nothing else shifts. A slightly less tidy transcript is much less noticeable than highlighting that runs two words ahead of the voice.

These fallbacks are counted. Each audio records in its provenance how many passages were not realigned, and a monitor raises an alert when more than one passage in ten fails over twenty-four hours, critical above four in ten. The absence of that check is exactly what let the September bug run for at least a week: every provenance record said so, and nothing was reading them.

A question worth asking any vendor

If you are choosing a speech tool for a website, this is invisible in an audio demo, so ask it directly:

  • Is the text under the player, and in the subtitles, the article's text or the text the voice received?
  • Does a lexicon correction show up on screen?
  • When alignment fails, does highlighting drift, or does it fall back cleanly?

You can test it yourself with one paragraph containing a number, a date, an acronym and a name fixed in the lexicon, watching the transcript while it plays. How we approach French reading in general, including what we deliberately do not do, is set out on our French text-to-speech page.

Give your articles a voice with WeDispatch

This blog is itself voiced by WeDispatch. Curious how it sounds on your content?

Book a demo

Read next