Why voices mispronounce words
A voice model learns pronunciation from the text and speech it was trained on. Common words come out right; names, invented words, brand names, technical terms and words borrowed from other languages are where it has to guess. Some English words also have two pronunciations depending on meaning, like read, lead, live, wind and tear, and the model has to pick one from context.
Six fixes, from simplest to strongest
- Generate again. Each take is a fresh attempt, and a word that wobbles once is often right the second time. Compare the new take with the current one and keep the better.
- Respell it phonetically. Write it the way it sounds, in ordinary spelling:
Shiv-awn,Hermy-own-ee,Kai-ruh. Use familiar words as building blocks;rowas in boat,cowas in moo. - Split it with hyphens or spaces.
Thessa-lonicaorThessa lonica. Splitting also helps long compound and technical words. - Move the stress. Spelling the stressed part more heavily often shifts the stress:
re-CORDdoesn't help, because the model reads in lower case, butrih-cordversusreck-erddoes. - Rephrase the sentence. For words with two readings, change the context so only one makes sense: "the lead singer" instead of "the lead", "she has read it" instead of "she read it".
- Swap the word. If a term won't come out right and the listener doesn't need that exact word, a synonym is quicker than a fight.
Names in other languages
A voice speaks in the language set for it, and pronounces names the way that language would. A French name in an English line gets an English reading unless you respell it. If a character speaks mostly in another language, set that voice's language to match; each voice has its own Language setting, so an English narrator and a French character can share a script.
Keep a pronunciation list
In a long project the same names come up again and again. Keep a short list of every respelling that worked, for example:
| Word | Write it as | Notes |
|---|---|---|
| Siobhan | Shiv-awn | Main character, every script |
| Kyrr | Keer | The city; one syllable |
| SQL | sequel | How the team says it |
Put it in the project's comments, where everyone the project is shared with can find it, and use it when writing new scripts, so names are right from the first take.
What respelling does to your files
The line's text is what goes into the download's Script.txt and Lines.csv, so a
respelled name appears respelled there. If you use those files for captions or subtitles, fix the spelling back
in the caption file. The audio is what matters for listeners; the captions are what matters for readers.
When a word gets dropped instead
If a word goes missing rather than coming out wrong, abcVoiceover usually catches it: every take is checked with speech recognition, and a piece that skips words is made again automatically. If one still slips through, generate the line again or split it into two shorter lines. More fixes are in the troubleshooting guide.