Teach dictation the fifteen words it keeps getting wrong

Last updated 11 September 2026

The diagnosis first

"Dictation isn't accurate enough" is almost never true as stated. Modern models handle ordinary English well. What is actually happening is that four or five words per message are wrong, every message, and they are always the same words.

They are your colleagues' names, your clients, your products, your industry's acronyms — words that are rare in the general English a model was trained on. The model is not failing; it has genuinely never encountered them.

Why a spoken form matters

There are two ways to handle a word a model gets wrong, and the difference decides whether it works.

"Siobhán" comes out as "sioban", "Shivon", "she von" and "sh von" depending on the sentence around it. Text replacement needs an entry for each. A spoken form needs one entry, because it is intervening at the point of recognition rather than patching the output.

Setting it up

1. Start with the names of people you write to Highest frequency, highest irritation. Colleagues first, then clients.
2. Say the word the way you actually say it Not the way it is spelt, and not carefully — the way it comes out mid-sentence at normal speed. If you say "Zscaler" as "zee scaler", that is the spoken form.
3. Type the exact spelling you want Including accents, capitals and hyphens. This is what will appear.
4. Add your acronyms These are the second biggest category. Decide the house form once — NHS or N.H.S., PO or P.O. — and be consistent.
5. Add product and company names Yours and your customers'. Particularly ones with unusual capitalisation.
6. Add one word a day for a fortnight, then stop The list converges fast. After two weeks you will be adding almost nothing.

What to add, in priority order

1. People you write to weekly. Especially names not native to your language.
2. Your company and product names. Particularly with odd capitalisation.
3. Client and supplier names.
4. Industry acronyms. Decide the house form once.
5. Technical terms. Kubernetes, nginx, Postgres, OAuth — rare in ordinary English.
6. Place names local to you. Models are trained on the whole language, not your county.
7. Words you use unusually. Every field has one — "instrument" to a lawyer, "charge" to a banker.

Things that will not work

• Adding five hundred words at once. Tempting, and counter-productive: a huge vocabulary makes the recogniser bias towards words you rarely say and introduces new errors. Add what you actually correct.
• Adding common words. Never teach it "there", "their" or "to". You will make it worse.
• Adding things you say once. A one-off name is faster to fix by hand.
• Expecting it to fix bad audio. Vocabulary corrects a model that heard you clearly and did not know the word. It cannot help if the microphone did not hear you — see fixing dictation mistakes.
• Expecting it to fix homophone choices. Whether you meant "principal" or "principle" is a judgement about meaning, and vocabulary does not do judgement.

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 use is quietly biasing the recogniser towards the wrong things.

It is also worth noting that this list is a mildly sensitive document — the names of everyone you work with, in one place. In Vaitly Voice it lives on your Mac with the rest of your settings and is not synced anywhere.