You add audio to your articles, press play on the first one, and the voice says "no fourteen" where the text says "No. 14". It sounds careless, and listeners notice it more than a typo on the page.
This post explains why it happens, why a smarter voice alone doesn't solve it, and what does, without asking your editors to write differently.
01Hear it first
For our home page we recorded one sentence twice, with the same voice. The first time the voice got the text exactly as written, and the second time Readerr prepared the text first.
"Bus No. 14 is back on High St. It carries approx. 3,400 riders a day, e.g. pupils on their way to school."
This is what the voice said each time:
- "No. 14" came out as "no fourteen". Prepared, it became "number fourteen".
- "High St." came out as "high seat". Prepared, it became "High Street".
- "approx." came out as "a prox". Prepared, it became "approximately".
- "3,400" came out as "three four hundred". Prepared, it became "three thousand four hundred".
- "e.g." came out as "e g". Prepared, it became "for example".
Nothing else changed between the two recordings. Hear both versions on the home page before you read on. It takes less than a minute.
02Why the voice guesses
A text-to-speech voice turns written characters into sound. It doesn't know it is reading a news article, and it can't ask what you meant. Printed news is full of short forms that people expand in their heads, and each of these five gives the voice a reason to guess wrong:
- "No. 14" The letters "No" are also the word "no", and the dot after them could just as well end a sentence. The voice read the word "no".
- "High St." "St." can stand for more than one word, and the voice has nothing to tell it which one. Here it picked neither and said something close to the letters.
- "approx." Nobody says "approx" out loud. The voice tries to pronounce the letters as if they were a word.
- "3,400" The comma splits the digits into two groups. The voice read the groups separately and lost the thousand.
- "e.g." Two letters with dots. The voice spelled them.
None of these is rare. A single local news story can hold all five. Every mistake pulls the listener out of the story for a moment.
03Why "the AI will figure it out" is not a fix
Voices keep getting better, and a newer one may read some of these correctly. But it is still guessing. The usual ways around the problem each have a catch:
- Waiting for a smarter voice. A guess can go one way in the second paragraph and another way in the fifth. It can change when the sentence around it changes.
- Writing numbers and abbreviations out by hand. Some voice vendors advise it. That works for a single recording, not for a newsroom that publishes all day and writes for readers first.
- Asking a language model to rewrite the article for listening. A model can also change what it wasn't asked to change. It might drop a word, turn one figure into another, or give a different result the next time.
Consistency matters as much as being right. The same abbreviation should sound the same in your fifth article and your five hundredth. For a news article, a figure read wrong is worse than a figure read awkwardly.
04Preparing the text with rules
Readerr takes a different route. Before the voice sees the article, the text goes through rules written for that job:
- a rule for "No." when a number follows it
- a rule for "St." at the end of a street name that ends the sentence, as in "High St."
- a rule for common short forms such as "approx." and "e.g."
- a rule that turns whole numbers such as "3,400" into words
Rules don't guess. The same input always gives the same output, in every article. That is why "3,400" is written out as "three thousand four hundred" before the voice reads it, every time.
The page itself doesn't change. Readers still see "No. 14" and "approx." exactly as your editor wrote them. Only the text the voice reads is prepared.
In some languages a number's ending changes with the word that follows it, so one figure can be said several ways in one article. A general guess handles this badly, because it has to understand the grammar of the whole sentence. Rules written for each language can pick the right form from the words around the number.
Rules have a limit too. A rule covers only what it was written for, and not every format has a rule yet. We add rules when we hear mistakes in real articles, which is also why the player tells listeners the audio is AI-generated and may contain errors.
Hear it on your own articles
7 days and 25,000 characters, about 5 articles. No card needed.
05Names need a human ear
Rules can't cover names of people, places and companies. There are too many of them, and new ones appear every day. For names, Readerr does use a language model, but only once per name:
- The model suggests how the voice should say a new name.
- The answer is saved.
- Every later article reuses it, so the name sounds the same each time.
Speech recognition won't catch a misread name. A common way to test audio is to turn it back into text and compare, which works for numbers and short forms. For names it hides the problem, because speech recognition writes a name back in its usual spelling even when the voice said it oddly.
If a name matters to you, listen to it.
06What your editors need to change
Nothing. They can keep writing "No. 14", "approx." and "3,400" the way your style guide says. Readers see what was written, and listeners hear words.
What does help is listening:
- Pick a few articles each week.
- Choose ones heavy with figures or unusual names.
- Play them from start to end.
You will hear the few things the rules don't cover yet, before your readers point them out. To hear the difference first, play both versions on the home page, then try it on your own articles with a 7-day free trial, no card needed.