How to Build a Custom Dictionary in Wispr Flow That Sticks
Dictation tools get about 95% of ordinary English right and then ruin the five words that matter. Your client’s surname. Your product name. The acronym your team invented in 2023. A custom dictionary fixes this, but only if you build it from real misses instead of guessing in advance.
How does the Wispr Flow dictionary work?
You add entries in Settings → Dictionary. Each entry maps how the word sounds when you say it to how it should be written. For most terms the spelling and the sound are the same, so you type it once. For unusual terms you add a phonetic hint — for example, a product called “Kynapse” that you pronounce “kin-apse” needs the hint or it comes out as “kinaptic”.
Entries are per-account, not per-device. Add one on your Mac and it applies on Windows within a sync.
How do you collect the words worth adding?
Do not build the list on day one. Spend a week dictating normally, and copy each miss into a running note.
- Dictate into a scratch file for a few days of real work.
- Whenever a word comes out wrong, paste the sentence into a text file called
misses.txt. - At the end of the week, sort the misses by frequency.
- Add the top 20 to the dictionary. Skip anything that appears once.
Twenty entries covers most professionals. Lists of 200 entries become their own maintenance problem, and long dictionary lists slow recognition as the model weighs more alternatives.
What kinds of words need phonetic hints?
| Word type | Example | Needs hint? |
|---|---|---|
| Common English name | Sarah, Michael | No |
| Brand with odd spelling | Canva, Zapier | Usually no |
| Made-up product | Kynapse | Yes |
| Technical acronym | CRDT, RAG | Sometimes |
| Non-English name | Nguyen, Siobhán | Yes |
Acronyms are their own trap. The model may spell a said acronym correctly and then lowercase it mid-sentence. If you dictate technical content, add both the acronym and the phrase it expands to.
Why does the dictionary sometimes not apply?
The most common cause is case. If your dictionary has Kynapse but the model outputs kynapse in lowercase, check whether the entry was saved with an explicit capitalisation rule. A second cause is that the rewrite step happens after recognition and can override a correct spelling — you will see this as a word that is right in the raw transcript and wrong in the final text.
When that happens, the fix is on the tone side, not the dictionary side: set the tone to a profile that says to preserve proper nouns, and test again.
How does this compare with a phone keyboard’s text replacement?
Phone replacements are literal find-and-replace on typed text; they cannot influence what a speech model hears. The dictionary sits earlier in the pipeline, so it actually changes recognition. That is why it works for names and not just for typo patterns.
Teams that dictate shared terminology — support reps, sales, clinical note-taking — get more from one shared list than from any individual tuning. Wispr Flow Dictation supports a single account-level dictionary, so agree on spellings before you add 40 variants of the same term.
If you prefer local processing, Superwhisper Dictation has a similar custom-vocabulary feature, plus per-mode vocabularies if you want to keep work and personal jargon separate.
Who should bother
Anyone whose dictated text contains names, product terms, or domain acronyms more than a few times a day — that is, most people dictating work email. If you only dictate grocery lists, skip it. The maintenance cost is real, and a 20-entry list needs pruning every couple of months as projects come and go.
Quick Summary
- The dictionary is account-level and syncs across Mac and Windows.
- Build it from a week of real misses, not guesses; 20 entries is usually enough.
- Made-up product names and non-English surnames need phonetic hints.
- If a word is right in raw and wrong in final text, the tone profile is overriding it.
- Phone text replacement cannot do this — it works after recognition, not before.
