Dictation is judged on the words it gets wrong, and the words it gets wrong are almost always names. Your product, your colleague, the library you use, the client you’re emailing. Every other word is fine and those three are wrong every time, which means every message needs a fix before you send it.
Why it happens
Speech models predict the most likely word for a sound. For “Siobhán” the most likely spelling, to a model that has never met her, is something else. For “VitaOS” it’s “Vita OS” or “vitals”. The model isn’t bad; it’s guessing without information you have.
The fix
Give it the information. Most dictation apps have a vocabulary or dictionary list. Add the exact spelling and the model is told to prefer it whenever the sound matches.
In Speaky Box it’s the Words page. Add a name once and both keys use it, in every app, so your product is spelt right in a Slack message and your customer’s name is right in a Claude prompt. There’s a hint field for tricky pronunciations if the sound and the spelling are far apart.
What to add on day one
- Your company and product names, including the weird capitalisation
- The people you message most
- Libraries, tools and file names if you code
- Clients and their products
- Any word you’ve corrected twice
Ten to twenty entries covers most people. Add more as you notice corrections.
What not to bother with
Ordinary words the model already knows. Adding “meeting” does nothing. Vocabulary is for proper nouns and jargon, and a list of five hundred common words can make things slightly worse, not better.
Five minutes on day one, and the correction habit disappears.