tl;dr
Custom vocabulary fixes the names and terms your dictation app keeps mishearing: enter the exact spelling once and it comes out right from then on. Build the list from evidence, meaning the terms you actually retyped in the past two weeks; fifteen to twenty-five entries covering colleagues, clients, products, and acronyms catches most repeat errors. In Auri Voice, add words under Settings > Dictation > Text > Vocabulary; the list works with every engine, and with Auri Cloud it also biases recognition itself (Pro). Auri Keyboard on iPhone keeps its own list per recognition profile.
Build the list from your own corrections
The best source for a custom word list is the editing you already did. Guessing produces entries that never fire and misses the surname you fix three times a week.
- Open your sent mail and Slack messages from the past two weeks.
- Note every term you remember retyping. Auri Voice's History (⌘3) helps here: search it for the wrong spellings the engine produced.
- Add the recurring ones, and stop at about twenty-five.
Two weeks of real messages is enough. Anything that matters shows up more than once in that window, and anything that appears once is not worth an entry.
Sort the list by how often you say the term rather than how wrong the transcript looked. A name you say daily that is slightly off costs more time than a spectacular failure you hit once a month.
What earns a slot
Custom vocabulary works because it raises the odds for words the model already considers unlikely. That means specific terms, not general ones.
Add these
- Names of people you write to or about weekly.
- Client, company, and team names.
- Product names, SKUs, repository names, internal tools.
- Acronyms and initialisms from your field.
- Place names and street names you type often.
Skip these
- Common words the engine already gets right.
- One-off names from a single meeting.
- Terms you type but never say out loud.
- Words you can paste from a source of record anyway.
Long identifiers are a poor fit either way. Account numbers and order codes are faster to paste, and no speech engine hears sixteen digits reliably.
Add the words in Auri Voice
The manual path is Settings > Dictation > Text > Vocabulary. Type the term into "Add a word, name, or phrase…" exactly as you want it written, casing and hyphens included. Every entry offers Teach by voice: say the term once so Auri learns how you sound, at the speed you actually talk.
The faster path starts from a mistake that already happened. Find the dictation in History, right-click it, and choose "Correct with Vocabulary…". Two fields appear, "When Auri writes" and "Always correct to"; fill them in and click Add Correction. Suppose your colleague is Siobhán and the engine keeps writing Shavon:
You said: "can you loop in Siobhán before the pricing call"
Auri typed, before the correction: "Can you loop in Shavon before the pricing call?"
Auri typed, after it: "Can you loop in Siobhán before the pricing call?"
The list works with every engine, because Auri applies entries to the transcript after recognition as exact replacements. With Auri Cloud, entries also bias recognition itself, so the right spelling is likelier on the first pass (Pro). Switching engines for noise or an accent costs you nothing here; the engine guide covers when a switch is worth it.
The list is also portable. The Vocabulary screen's overflow menu has Import… and Export…, and the file is JSON, so a teammate can inherit your twenty correct spellings instead of repeating your two weeks of fixes.
Custom words in Auri Keyboard on iPhone
Auri Keyboard keeps two vocabulary lists, one per recognition profile. For keyboard dictation: Auri's Settings > Keyboard Dictation > Recognition > Custom Vocabulary, turn on Use Custom Vocabulary, then Add Words. Recordings, imported audio, and Smart Notes read a separate list in the same spot under Transcription & Smart Notes > Recognition.
Engine choice matters on iPhone in a way it does not on Mac. Ultrafast (up to 100 terms), Supernova, Legacy, and Azura read the list; Wizper and On-Device do not. If names drive your corrections, pick one of the first four in the same Recognition screen.
Spelling and how you actually say it
An entry works when it matches what leaves your mouth. Most disappointing lists fail here rather than in the feature itself.
| Term type | How you say it | What to enter |
|---|---|---|
| Acronym said as letters | Separate letters | The letter form you use in writing |
| Acronym said as a word | One spoken word | The written form; say it as a word when you teach it |
| Name spelled unlike its sound | Your usual pronunciation | The correct spelling, taught by voice |
| Hyphenated product name | Two words | The exact written form with the hyphen |
Casing and hyphens are worth being fussy about, since the whole point is not fixing them later. Enter the form you would ship in an email to a client, then use Teach by voice at your quick everyday pace. The careful version you say while testing is not the version you will use at five o’clock.
Test in a way that proves it worked
Adding entries feels productive, so it is easy to skip verification and keep a list full of things that do nothing. Ten minutes of testing fixes that.
- Write one sentence that uses five of your new terms naturally.
- Dictate it before adding the entries and count the errors.
- Add the entries, dictate the same sentence, and count again.
- Dictate one real email and one real Slack update to confirm it holds.
- Remove entries that never fired.
Test at your normal speaking pace. Terms pronounced carefully during a test and quickly in real work behave differently, and the real work is what you are tuning for. Keep the count: going from six corrections to one on the same sentence tells you to stop tuning and get back to writing.
Maintenance that takes two minutes a month
Vocabulary decays: clients end, projects close, and the names you needed in March are noise by September.
- Add three to five terms when a project starts.
- Remove terms from finished work once a quarter.
- Export the list (overflow menu > Export…) before a reinstall or a new Mac.
- Share the exported JSON with a teammate so the spellings are not stuck on one machine.
- Update the entry when someone changes their name. Getting a colleague’s new surname wrong because the list still holds the old one is worse than a generic error.
On iPhone, recheck the list after an engine change, since Wizper and On-Device ignore it.
When vocabulary is the wrong fix
A word list solves a narrow problem. Reaching for it when something else is broken wastes an afternoon.
| Symptom | Real problem | Fix |
|---|---|---|
| Ordinary words fail too | Mic distance or noise | Get closer, quieter room |
| Everything looks phonetically wrong | Language mismatch | Set the recognition language |
| Nothing appears at all | Permissions or focus | Grant Accessibility, retest per app |
| Rare terms fail after a good list | Engine fit | Compare another engine on the same script |
| Wording changed but words were right | The cleanup pass | Judge cleanup separately |
That last row causes real confusion. Auri Voice's cleanup pass removes filler, fixes punctuation, and applies light formatting on purpose, so changed wording is expected behavior rather than recognition drift.
And when a term keeps coming out split or mangled no matter what you enter, correct the output instead of the recognition: add a pair with When Auri writes set to the wrong form and Always correct to set to the right one. The replacement runs on every engine, so the mishearing stops mattering.
FAQ
How many custom words should I add?
Fifteen to twenty-five to start. A short list drawn from real corrections beats a long list of terms you might say someday, and it is easier to notice which entries are earning their place.
Should I enter the spelling or the pronunciation?
Enter the correct spelling, exactly as you want it in writing, then use Teach by voice in Auri Voice to say the term once so Auri learns how you sound. Acronyms matter here: a term you say as separate letters behaves differently from one you say as a word.
Which engines use custom vocabulary?
In Auri Voice, every Mac engine applies your list, and Auri Cloud also biases recognition itself (Pro). In Auri Keyboard, Ultrafast, Supernova, Legacy, and Azura read the list; Wizper and On-Device do not.
Does Apple Dictation support custom vocabulary?
Its personal vocabulary options are limited compared with dedicated dictation apps. If names are your main source of corrections, that alone is a reason to use an app built for it.
Does iOS Text Replacement fix dictation errors?
Text Replacement expands typed shortcuts, so it is not a recognition fix. It can help with long terms you type often, but it will not make the speech model hear a name it does not know.
My list is not helping. What now?
Check whether the failures are actually vocabulary. If ordinary words fail too, the problem is mic distance or language settings. If one term still fails after a good list, add a correction pair from History so the wrong output is replaced either way.