Productivity
Which twenty words are worth teaching it
Teaching a dictation app vocabulary is a fifteen-minute job. Choosing the wrong twenty words makes accuracy worse. Here is how to pick, by what you actually do.
Last updated
The three conditions
Every entry should satisfy all three. Most people's instinct is to add everything specialised, which fails the first condition and quietly degrades the model's output.
- Frequent. You say it at least weekly. Something said twice a year is faster to fix by hand than to maintain.
- Rare in ordinary English. This is why the model gets it wrong — it genuinely has not seen it. Proper nouns, jargon, product names, acronyms.
- Consistently wrong. Check first. Modern models handle a lot of technical vocabulary correctly and an entry for something already right is pure cost.
The twenty, by profession
Software
Kubernetes, nginx, Postgres, Redis, Kafka, OAuth, JSON, YAML, kubectl, Grafana, your service names, your repository names, your framework, your cloud provider's product names, and the four colleagues whose surnames nothing spells right. Almost all of these are mangled by default because they are rare outside engineering.
Law
Counsel's names, chambers, the courts you appear in, your clients, statutes you cite constantly, and the house form for your acronyms. Legal English also contains ordinary words used unusually — "instrument", "consideration", "without prejudice" — which mostly come out right and are worth checking rather than assuming.
Medicine and healthcare admin
Drug names, procedure names, the departments and the consultants you refer to, and the local trust or practice names. Drug names in particular are both high-frequency and high-consequence, and are exactly what a general model has not seen.
Accountancy and finance
HMRC, VAT, CT600, SA100, PAYE, UTR, your software's names and your clients' company names. Decide the house form once — HMRC or H.M.R.C. — and teach it that, which also settles an inconsistency that runs through most practices.
Recruitment and sales
Client company names, your product names, your competitors, and the job titles your sector uses. Candidate and prospect names are disproportionately international, which is precisely where general models are weakest.
Everyone
Your own company and product names, the five people you write to most, and any place name local to you. Models are trained on a language, not on your county, and local place names are a reliable source of low-grade irritation.
What not to add
| Do not add | Why |
|---|---|
| Common English words | You will make it worse. Never teach it "there", "their" or "to" |
| Anything already correct | Pure cost, no benefit. Check before adding |
| One-off names | Faster to fix by hand than to maintain |
| Whole phrases | That is a snippet, not a vocabulary entry |
| Homophone preferences | Whether you meant "principal" or "principle" is a judgement about meaning |
| Hundreds of terms at once | Rare words start competing with common ones |
Finding your actual twenty
- Dictate normally for one week Do not try to design the list in advance. You will guess wrong.
- Keep a note of every word you correct In the same document you are dictating into is fine. It takes two seconds each time.
- At the end of the week, count You will have about fifteen items and be surprised how short it is. The same names, over and over.
- Add them, with the spoken form you actually use Not the careful pronunciation — the one that comes out mid-sentence at normal speed.
- Add one or two a day for another fortnight, then stop The list converges fast. After that you are adding noise.
Keeping it useful
Review it twice a year. Clients leave, projects end, the colleague whose surname you taught it has moved on. A vocabulary full of words you no longer say is quietly biasing the recogniser towards the wrong things, and it is the kind of decay nobody notices because each individual entry seems harmless.
It is also worth remembering what this list is: the names of everyone you work with and everything you work on, in one file. In Vaitly Voice it stays on your Mac with the rest of your settings and is not synced anywhere.
Questions
How many words should I teach it?
Fifteen to thirty. That covers the recurring corrections for almost everyone. Hundreds of entries makes accuracy worse, because rare words start competing with the common ones you actually say.
Will it handle drug names and medical terms?
Not by default — they are rare in general English. Teaching the twenty you use daily fixes it, and it is worth doing carefully because a misrecognised drug name is not a typo.
Can I import a glossary?
In VV, entries are added by hand, which is a deliberate constraint: a bulk import is how a vocabulary becomes five hundred entries and starts making accuracy worse. Twenty words takes fifteen minutes.
What is the difference between this and a snippet?
Vocabulary gets one word spelt correctly. A snippet turns a spoken trigger into a whole block of text. You want both, and they solve different problems.