How to Add Words to Mac Dictation That It Gets Wrong

Apple’s standard Mac Dictation guide does not document a personal word-list editor. Voice Control has a separate custom vocabulary for supported languages. In Ducduc, use Personal Dictionary for a spelling hint, learned corrections for eligible edits, and Exact replacements only when you want a deliberate substitution.

The right fix depends on the error. A name that needs context is different from a phrase you always want replaced in one app. None of Ducduc’s tools retrains Apple’s acoustic model. This guide is written by Ducduc’s maker and describes the reviewed build, with illustrative examples rather than measured accuracy results.

If you want to keep using Apple’s built-in tools

Start by identifying which feature you use:

Feature	Where to look	What the documented vocabulary applies to
Standard Dictation	System Settings → Keyboard → Dictation	Apple does not document a user-managed general vocabulary list here
Voice Control	System Settings → Accessibility → Voice Control → Vocabulary	Voice Control, in supported languages
Ducduc	Menu-bar Dictionary, or Cleanup settings → Manage Personal Dictionary	Ducduc’s own recognition/cleanup workflow

For Voice Control, open Vocabulary, add the word or phrase, and follow the prompts. If Vocabulary is dimmed, Apple says the selected language does not support it. Importing a term list is also supported. Use Apple’s current vocabulary instructions for your macOS version.

Voice Control is a hands-free accessibility system, not just a dictionary toggle. When it is on, Apple says standard Dictation is unavailable. Do not assume that adding a Voice Control term teaches keyboard-triggered Dictation or a third-party app.

Apple’s Siri and Dictation privacy notice describes contact names as personalization information. A real colleague’s Contacts entry may help with their name, but this is not a documented general dictionary for code identifiers or project codenames.

Diagnose one recurring error before adding rules

1. Dictate a full sentence containing the term into Notes.
2. Open Dictation History and compare raw and final text. If raw spelling is correct but final spelling changes, start with cleanup and existing rules.
3. Write down the wrong phrase and intended spelling. Include enough neighboring words to distinguish a genuine correction from a change of mind.
4. Choose the weakest tool that solves the problem, then test both a positive example and a sentence that must remain unchanged.

For English inside Chinese, use the mixed-language checklist. If there is no transcript at all, adding words will not fix missing speech assets or unavailable recognition.

Add vocabulary when context should decide

Open Dictionary from the menu bar, or Settings → Cleanup → Manage Personal Dictionary…. Type the desired spelling in New word and press Return. Add terms you actually say, one at a time. The import button accepts Ducduc backups, not arbitrary exports from another app.

Ducduc supplies known spellings to cleanup, and the classic recognition path can receive vocabulary hints. Its built-in lexicon already includes common developer vocabulary. Apple documents contextual strings as recognition hints; that does not establish acoustic retraining or identical support in every speech engine.

A term is a candidate, not a forced mapping. Adding Vercel does not justify changing every vaguely similar word in a deployment conversation. If the term was omitted entirely, a text-only model must not invent it simply because it appears in your dictionary.

Use an exact replacement for a deliberate mapping

Under Exact replacements in the dictionary window:

1. Enter the recurring wrong phrase under Replace this word or phrase.
2. Enter the intended spelling under With this spelling.
3. Choose All apps or the relevant app under Applies in.
4. Click Apply replacement, then retest.

Illustrative rule: a recurring right the migration note → write the migration note mapping scoped to your coding app. Test “Please right the migration note” to confirm the mapping, and “Keep the panel on the right side” to confirm unrelated prose survives. These are constructed examples, not actual recordings.

Avoid global right → write: both are real words. Prefer a distinctive phrase; Chinese has no spaces between words, so broad single-character rules are particularly risky. An app-specific rule takes precedence over a global rule for the same source phrase.

The rules run locally before AI cleanup, including when cleanup is unavailable. They protect larger Latin words, identifiers, URLs, email addresses and inline code. They do not cascade: with A → B and B → C, an original A becomes B. A rule is still a deliberate substitution, so review the final text. Lowering cleanup strength does not disable it.

Why fixing it once may not teach the app

After insertion, Ducduc can observe eligible edits for a short period through macOS Accessibility. A readable target field is required; some terminals, browser fields and Electron editors expose too little information. The dictionary’s Learned corrections list shows what was actually saved.

The reviewed development checkout can also use repeated, sound-related edits in matching app and neighboring vocabulary to learn a context-specific correction. It requires repeated evidence and blocks conflicting destinations. This does not mean an ordinary word or Chinese homophone becomes a global replacement. Unambiguous malformed-English corrections have a separate local path.

Small spoken replacements through selected-text voice editing can provide another learning route when eligible. Style changes, translation, changed quantities, code and wholesale rewrites are not pronunciation evidence. A user edit can also be a change of mind; these filters reduce ambiguity without proving how the audio sounded.

Inspect learned pairs, delete an incorrect one and undo deletion if necessary. A Replace All offer can help with other occurrences; if a field cannot be written through Accessibility, use Copy fixed text when offered.

Back up, retest and remove stale rules

Under Backup, Export… saves a Ducduc JSON backup. Import… merges supported backups with existing entries, including exact-rule app scopes. Keep this file private if it contains names or internal project terms. Clearing Dictation History also clears learned corrections; deliberate exact rules are preserved and must be removed separately.

Learned neighboring-word context stays local. Relevant vocabulary and Latin correction hints may be supplied with cloud cleanup; the data-flow guide explains the privacy boundary. A .local model address is not necessarily the same Mac.

Try Ducduc with one recurring term and one control sentence. If it still fails, report the intended spelling, raw/final pair, target app and build, and whether a hint, learned pair or exact rule existed. For a script preference rather than a misheard word, use Traditional Chinese output settings.