Changing the voice of an already-generated audio: what happens, and what it costs

You've added audio to an article, you've listened to it, and the voice isn't right: too deep, too flat, the wrong register for the content. The next question comes straight away: can you change the voice without redoing everything, and what does it cost? The honest answer rests on one distinction, and it's worth understanding before you press a button.

Changing the voice means rebuilding the audio

Let's be plain about it: changing the voice of an already-produced article isn't an edit, it's a fresh build. A synthetic voice can't be "replayed" the way you'd shift the tone of an existing recording. Each voice reads the text in its own way, with its own timbre, pace and liaisons. Asking for another voice means asking the engine to read the whole article again, from the first word to the last.

In practice, the tool takes you back to the voices screen, you pick one, and you relaunch the creation. On the server side there's nothing special: it's a generation like the first. The file produced is a new file, and it replaces the previous one. Which also means, in passing, that the result won't be a comma-for-comma twin of the old one: a rebuild is a fresh reading, not a copy. If the voice pleases you, that doesn't matter; you just need to know it.

What it charges

Since it's a full build, it's billed as a full build. A credit is worth a thousand characters, and a generation charges for the number of thousands of characters it starts into, never less than one credit. A thousand-word article runs to around six thousand characters, so re-voicing it with a different voice costs about six credits: exactly what the first version cost. Changing the voice a third time would cost six more. There's no half-price because you've "already paid": the engine redoes all the work, and the work is the same.

Two safeguards hold, and they matter. First, nothing is charged without your explicit go-ahead: the amount is shown before you launch, you see it and you decide. Second, a build that fails costs you nothing; if an audio stops partway, no credit is taken, and you can relaunch. If you're unsure what an audio will cost before you commit, we set out the maths in estimating an audio's cost in credits.

When you should definitely not change the voice

Here's the most common trap, and the most expensive: you hear a single mispronounced word, a mangled proper noun, and the reflex is to regenerate everything in the belief it'll "set things straight". That's the wrong tool. A wrong word isn't fixed by changing the voice (the new voice will often slip in the same place, for the same reason), and it doesn't deserve paying for the whole article again.

For that case there's targeted correction: the tool re-synthesises only the words you point to, with the same voice, and charges only for those characters, to the hundredth of a credit, without starting a whole new credit. Fixing a misread "Ploërmel" is a handful of characters, not an article. We describe the mechanism in detail in fixing a sentence without regenerating everything. The simple rule: change the voice for an editorial choice across the whole article; use a targeted correction for one word that sounds wrong.

Choosing the voice well the first time

The best way not to pay twice is to listen before you produce. The tool has the chosen voice read the start of your real article, and it's that short sample that decides, far better than a demo on a generic text: a voice that holds up on a showroom sentence can disappoint on your own style. If it doesn't fit, you change the voice right there, in the sample, before any spend. For what to judge a voice on in a few minutes, our method for evaluating a neural voice sets out the points to listen for. And the choice of voice by content is far from neutral: it's one of the settings that separates an audio you tolerate from one you actually listen to, at the heart of what it means to read French well.

What to take away

Changing the voice of an already-generated audio is possible, it's simple, and it's a full rebuild: the price of a creation, a new file, and never without your go-ahead. Keep that move for a real editorial decision about the whole article. For a single misread word, targeted correction costs pennies of a credit and keeps your voice. And to avoid choosing blind, listen to the sample before you launch: a voice is better picked at the start than corrected after.

Read next

• How long does an article's audio actually last?
• Reading a list aloud: what a bulleted list becomes to the ear
• Adding an audio version to your Kajabi content