Teach dictation your vocabulary: names, jargon and acronyms

Why dictation gets your words wrong

The speech model hears a sound and writes the most likely words for it. For common words that works well. For your words it often does not: without any help, the standard model Deixis ships writes shadcn as "shadkin" and useEffect as "use effect". The same can happen to a colleague's surname or a product your company sells.

It is tempting to hope an AI tidying pass will catch these. It will not. Deixis's optional cleanup step rewrites prose, not hearing: it never learns what you actually said, so a misheard name stays misheard. Getting names right is the dictionary's job.

How the dictionary works

The Dictionary tab has four parts — Vocabulary libraries, Words to expect, Fixes and Filler words — and on Windows it is the same thing as the file %USERPROFILE%\.deixis\dictionary.toml, so you can edit either.

Words to expect are handed to the speech model as a hint of what it is likely to hear, which steers it toward those words when it hears something close to one. This is the lever that changes what the model writes in the first place.

Fixes are find-and-replace, applied to the text after recognition. If it keeps writing "shadkin", a fix turns "shadkin" into "shadcn" every time. A fix matches whole words only, ignores capitals when it matches, and lands the replacement exactly as you typed it. Fixes run in the order they are listed.

Filler words are stripped from what you say. "Um", "uh" and "erm" are already on the built-in list; add your own habits, such as "basically".

In the file, all three look like this:

vocabulary = ["Siobhan", "Okonkwo", "EBITDA"]

extra_fillers = ["basically"]

[[replacements]]
from = "shadkin"
to = "shadcn"

On Windows, adding a word takes one shortcut. You do not have to open anything. Select it wherever you see it — an email signature, a contract, a page of your notes — and press Ctrl+Win+D. The word or short phrase goes straight into your vocabulary. With nothing selected, the same key opens a small popup where you type the word instead. You can also hold Ctrl+Win and press D from the mode palette; the gestures page lists every key.

Why the list stops at 36

Only the first 36 terms reach the speech model. Each term adds a little to the time it takes to decode what you said, and past 36 the accuracy measurably gets worse — so the cap is a measurement, not a guess. The editor tells you when you are over.

Words past the cap are not thrown away. Up to 256 of them still go to the cleanup model, if you use one, which also carries your spellings into translation. But they stop steering the recognition itself.

Fixes have no cap and cost nothing. Once your words to expect are full, put new corrections in the fixes list instead.

Vocabulary packs

For whole fields rather than single words, Deixis ships five vocabulary packs: Software development, Accounting & finance, Business & product, Medical and Legal. Each is sized to fit under the 36-term cap on its own. Switching on several overflows the cap, and the editor says so rather than hiding it.

The Software development pack is on by default. If you do not write code, switching it off leaves its room under the cap for your own words. You can also write your own pack: on Windows, drop a .toml file into %USERPROFILE%\.deixis\packs\ and it appears beside the built-in ones.

Which words to add

• Start with words it has already got wrong. Loading every term you know wastes the 36 places on words it would have heard anyway.
• Names first. Colleagues, clients, authors you cite, characters, places, products — proper nouns are the words a general model is least likely to know.
• A consistent mistake wants a fix. If it writes the same wrong thing every time, a fix repairs it without using one of the 36 places.
• A word it has never heard wants a vocabulary entry. If the mistakes vary, the model does not know the word, and telling it to expect the word is what helps.
• Acronyms and odd capitals suit fixes. The replacement lands exactly as typed, and Deixis's cleanup is not allowed to re-case anything but the first character, so your spelling survives.
• Check the packs before building a list. If you work in finance, law or medicine, a pack may already carry the terms you were about to type in.

If the dictionary is not enough, two other levers help. Settings → Speech model offers a larger model that is more accurate and slower to decode, and the control says by how much. And no software recovers a word the microphone never caught, so a headset beats a laptop's built-in microphone in a room with other people in it. The dictionary documentation covers both.

On your phone, and what leaves your machine

On Android, the dictionary is one of the features carried over to the phone, along with translate, the Library and both speech models. The platforms page lists what the Mac app and the iPhone keyboard, in beta on TestFlight, do today.

The dictionary is a file on your own machine, and it is free, with no account: the dictionary, all five packs and your own are on the list of things that are never gated. Recognition runs on your device, so your words steer it without going anywhere. If you have set up cloud cleanup, your vocabulary terms are sent along with any text it tidies or translates, so it keeps your spellings; with no endpoint, nothing leaves. The full list is on what leaves your machine.

To start, download Deixis, dictate a paragraph you write often — an email to a regular client, say, as in dictation for Outlook — and add whatever it gets wrong. The free voice to text page covers everything else that costs nothing.