Why is the same word misheard every time?
Usually the word is new to the model, which Vocabulary fixes, or the model is too small for your accent or the noise around you, which a larger model fixes.
A word that is wrong once might be noise. A word that is wrong every single time has one of two causes, and they have different fixes.
#Cause 1: the model has never heard this word
Product names, people’s names and specialised jargon are the most common case. The model writes the closest ordinary word it knows.
- Open Vocabulary.
- Add the word exactly as it should be spelled.
- Dictate the same sentence again.
The word is now recognised correctly, or close enough for the next fix below.
If it still comes out close but wrong, turn on AI cleanup, which corrects near misses against your Vocabulary list.
#Cause 2: the model itself is struggling
If the word is ordinary English and still wrong every time, the cause is usually the model, not the word. This shows up most with a strong accent, background noise, or the Tiny model.
- Open AI Models.
- Switch to Small, which handles accents and noise better than Base or Tiny.
If the word is unusual (a name, a brand, a technical term), start with Vocabulary. If ordinary words around it are also getting mangled, start with the model.
Neither fix is magic. Vocabulary raises the odds of a correct match; it does not guarantee one. A larger model is more accurate in general, not on every word every time.