How to add custom words to your dictation

Say your colleague’s name into any dictation tool and watch what comes out: a phonetic guess, wrong in a way autocorrect won’t fix, because the model has never seen the word. Names, product codenames, internal jargon, that one client with the unpronounceable domain — this is where dictation loses people, and it’s fixable. Two tools do the teaching: one ships with macOS, one ships with Snapbak.

Why dictation gets names wrong

Speech models transcribe from everything they’ve heard — public text, common words, common names. Your manager’s surname, your startup’s name, the npm package you keep mentioning: statistically rare, phonetically ambiguous, and therefore the first casualties. It isn’t a bug you can report; it’s a boundary of the model’s world. The fix is always the same shape: tell the tool what the word is, once, so it stops guessing.

macOS text replacements: the system‑wide fix

System Settings → Keyboard → Text Replacements lets you map a shortcut to an expansion: type ssx, get Saksham Sharma Systems, Inc. It works in most text fields, it syncs over iCloud, and it has one trait worth knowing: it’s a spelling fix, not a listening fix. Dictation still hears what it hears; the replacement fires when the typed result matches your shortcut. So teach it the misheard form: when dictation writes a name wrong — “Sachem Sharma” instead of the real one — map that exact wrong string to the right name, and every future dictation self‑corrects.

The catch is volume. Twenty replacements is a weekend project; two hundred is a hobby — and the mapping only fires on exact matches of what the model chose to write, which drifts as models update.

A dictation dictionary that learns the word

Snapbak’s dictation has a custom dictionary built for exactly this: you add the word — a name, “Kubernetes,” “SOC 2,” a client’s product — and the speech model itself recognizes it in speech from then on, in any app you dictate into. Because it’s a listening fix rather than a spelling fix, the word comes out right the first time, no shortcut to remember, and the dictionary lives on the device with everything else. It’s English‑only dictation on the Neural Engine — the accuracy numbers are published, and the architecture post explains why the vocabulary was the hard part.

There’s a dictation history on the device too, which doubles as the to‑teach list: scan what you dictated last week, and every word the model mangled is a candidate for the dictionary. Add it once, and it stays added — model updates don’t undo your vocabulary.

Which to use

Both, for different jobs. Text replacements handle formatted expansions — email addresses, addresses, boilerplate — anywhere on the Mac. The dictionary handles the words dictation actually mishears, everywhere you dictate. The offline dictation overview covers the privacy side of the trade; the short version is that both approaches keep the words on your machine.

Snapbak is free for the first 100 users: download it, add the three names you say most, and hear the difference on the next sentence.