How to Make Dictation Spell Names, Brands and Jargon Correctly

Dictation gets the hard part right and the easy part wrong. It will handle a long, complicated sentence perfectly, and then spell your colleague’s surname three different ways in the same paragraph.

If you write about the same people, products or technical terms every day, this is the thing that makes dictation feel unreliable — not the transcription quality.

Why it happens

A speech model does not hear letters. It predicts the most likely sequence of words for a given sound, and “most likely” is decided by how often those words appear in ordinary language.

That is exactly why proper nouns lose. Your surname, your company’s product name, a piece of industry jargon — all of them are rare. A common word that sounds roughly the same is far more probable, so the model picks it. The output is confident, fluent and wrong.

Worse, it is wrong consistently. The same name gets the same wrong spelling every time, so the correction becomes a daily chore. For anyone dictating to avoid typing, retyping the same word twenty times a week defeats the purpose.

The three things that actually help

1. Give the model a vocabulary hint. Most speech APIs accept a short prompt alongside the audio — a handful of correctly spelled words that tell the model what to expect. Providing the words you use often measurably improves the odds that they come out right the first time.

2. Keep a personal replacement list. For the cases the model still gets wrong, a simple substitution — “this misheard form always means that correct word” — fixes it instantly and permanently. It is unglamorous and it works.

3. Catch the mishearings you have not seen yet. This is the part most tools skip. A name has more than one way of being misheard, and a fixed replacement list only catches the variants you have already run into and typed in. Matching on how a word sounds rather than how it is spelled catches new variants of a word you have already taught it.

That third step has an obvious danger: an over-eager phonetic match will “fix” real words that merely resemble your term, and mangling ordinary English is far worse than misspelling one name. The protection is a guard list of everyday words that must never be replaced — which has to be per-language, because what is an everyday word in English is not one in Hungarian.

What this looks like in practice

In BeltoVox this is the Word Corrections feature, and all three parts are there:

• You list the words you care about, along with the misheard forms you have already seen.
• The correct spellings are sent to the speech model as a vocabulary hint, so many mistakes never happen in the first place.
• An optional phonetic pass catches new mishearings that sound like a word on your list, guarded by an editable everyday-words list for each of the 51 supported dictation languages so ordinary vocabulary is left alone.

All of it stays on your machine. The word list is a local file — nothing about your vocabulary is uploaded, which matters when the words in question are client names or unreleased products.

The habit worth building

Do not try to fill the list up front. Add words when they annoy you: the first time you correct a name by hand, add it. After a week or two of that, the list covers what you actually say, and dictation stops feeling unreliable — because the thing that made it feel unreliable was never the transcription. It was the same wrong name, every single day.

BeltoVox is a Windows dictation app you buy once at $59, with a lifetime licence. It types into whatever application you are in, cleans up or rewrites what you said, and keeps your personal vocabulary on your own machine.