Fixing one sentence in already-generated audio, without redoing it all
A typo, a mispronounced name, a last-minute update: how to redo a single passage of a narrated article without regenerating the whole thing, what it costs exactly, and why the cut is inaudible.
In classic audio, the smallest fix means redoing everything. A mispronounced name, a wrong figure, a detail that lands after publication: reworking the passage means re-recording the whole article, or running a full new synthesis. It is slow, it is billed at the price of a fresh article, and the outcome is familiar: more often than not, the mistake just stays on air because redoing it is not worth it. A newsroom that publishes fast ends up living with its audio typos.
The problem is not voice quality, it is the unit of work. As long as the unit is "the article", every fix costs an article. Targeted correction changes the unit: it reworks only the passage you rewrite, and leaves the rest bit-for-bit intact.
The mechanism, described precisely enough to be checked
The starting point is a property few tools expose: the generated audio is timestamped word by word. Every word in the transcript knows its exact position in the file. That is what makes the rest possible, and it is the same foundation as synced highlighting or sharing a clip at a precise second.
The fix takes three moves. First, you select the passage to correct in the transcript by clicking the first and last word; thanks to the timestamps, the tool grabs the exact audio portion they correspond to. Then a field pre-fills with the current text of that passage: you correct it, and only that passage is resynthesised, in exactly the same voice and settings as the rest of the article. Finally, the new segment is dropped back in place of the old one, on a silent boundary in the voice, down to the frame. To the ear there is no click, no seam, no shift in timbre: the rest of the audio has not moved, because it was not regenerated.
This "recompute only what changes" logic governs the whole preparation of an article, described on the French text to speech page. A correction is just a special case: the smallest possible change, treated as such.
What it costs, with no flattering rounding
Speech synthesis is billed per character, and a correction is no exception, but it pays only for its own. The rule is simple: a thousand characters make one credit. Reworking a twenty-character name therefore costs two hundredths of a credit, and correcting ten words in a thousand-word article costs the equivalent of those ten words, not the price of the article. It is never free, and it is never the price of a full generation.
Two safeguards frame the spend. The cost is shown while you type, before anything is debited, rounded up to the hundredth, and it asks for explicit confirmation: no script can spend on your behalf. And the old version keeps being served until the new one is ready, so a correction in progress never leaves an article silent. For comparison, a full regeneration is billed like a generation: one credit per started thousand characters. The full cost breakdown of a whole article is in how much article audio costs.
When to correct, and when to regenerate
Both moves coexist, and the right one depends on the size of the change. Targeted correction is made for retouching: a typo, a mangled proper noun, a misread acronym, a figure that changes, a reworded sentence. It is ready in seconds, with no queue, which matters for a wire story or breaking news fixed within the minute.
Regeneration remains the right answer when the article has been deeply rewritten, when you change the voice, or when the structure has shifted so much that stitching would no longer make sense. The case of an article corrected after publication on a WordPress site, and what triggers one path or the other, is covered in regenerating the audio of an updated article.
One last reflex saves you from fixing the same thing twice. If a name, a brand or a phrase recurs across several of your pieces and reads badly everywhere, one-off retouching is the wrong scale: the pronunciation lexicon sets the expected pronunciation once, and it applies to future articles without you going back over them.
What it changes in a newsroom
For a team that publishes fast and fixes often, the point is not only cost, it is the decision threshold. When reworking a mistake costs an article and takes as long as a regeneration, you let it slide. When it costs two hundredths of a credit and takes seconds, you fix it. Targeted correction also fits the pre-publication review loop: a reviewer can rework a doubtful passage before approving, as described in reviewing audio before publishing, and scheduled publishing is respected, the approved audio only going out at its set time.
We prefer to put it this way rather than promise a machine that never gets it wrong. It does get things wrong sometimes, as any reading does; what matters is that the fix is fast enough and cheap enough that you actually make it, instead of living with the mistake.
Give your articles a voice with WeDispatch
This blog is itself voiced by WeDispatch. Curious how it sounds on your content?
Book a demo