Questions
Can dictation learn your names? Some can
Apple's built-in dictation cannot. Apps with a custom vocabulary can, and it is the single change that turns dictation from irritating to reliable.
Last updated
Why names are the problem
A speech model learnt from an enormous amount of general English. Your colleague Siobhán, your client Zscaler and your industry's acronyms are not in general English, so the model has genuinely never encountered them. It is not failing; it is guessing, from a vocabulary that does not contain the answer.
That is why dictation feels inaccurate to people whose work is full of proper nouns and fine for people whose is not. The general accuracy is the same; the proportion of unknown words is completely different.
What "learning" means here
Not that the model retrains. The weights are fixed when the app ships and do not change from use.
What happens is that a layer around the model is told about your words. A vocabulary entry pairs a spoken form — how the word sounds when you say it — with the spelling you want, and the recogniser is biased towards that spelling when it hears that sound.
| Text replacement | Vocabulary with a spoken form | |
|---|---|---|
| When it acts | After the mistake | Before it |
| What it needs | The exact wrong string | How the word sounds |
| Handles variations | No — one rule per variation | Yes — one entry |
| Available in macOS dictation | Yes, as text replacement | No |
| Fixes a misheard name | Sometimes, by luck | Yes |
That middle row is why this matters. "Siobhán" comes out as "sioban", "Shivon", "she von" and "sh von" depending on the sentence around it. Replacement needs a rule for each; one spoken form covers them all.
Untaught
ask sioban to send the deck to N.H.S. digital and check if key ran has the P.O. from Zed scaler
After fifteen entries
Ask Siobhán to send the deck to NHS Digital and check if Kiéran has the PO from Zscaler.
How long it takes
Fifteen minutes, once, for about fifteen words. Almost everyone expects a much longer list and is surprised by how short theirs is — the same names recur, which is exactly what makes them worth teaching.
- Dictate normally for a week Note every word you correct. Do not try to design the list in advance.
- Count About fifteen items, and they will all be proper nouns.
- Add each with the spoken form you actually use Not the careful pronunciation — the one that comes out mid-sentence at speed.
- Stop Adding hundreds makes accuracy worse, because rare words start competing with common ones.
What it cannot do
- Choose between homophones by meaning. Whether you meant "principal" or "principle" is a judgement, and vocabulary does not do judgement.
- Fix bad audio. If the microphone did not hear you, there is nothing to bias. See fixing dictation mistakes.
- Learn from your corrections automatically. In most apps you add entries deliberately. That is slower and much more predictable than an app quietly deciding what you meant.
- Cope with hundreds of entries. More is worse past about thirty.
Which apps have it
Vaitly Voice does, and most paid Mac dictation apps do in some form. macOS's built-in dictation does not — text replacement is the nearest thing and it is not equivalent.
If you are choosing an app and this matters to you, check specifically whether entries carry a spoken form or are simply text substitutions. The two are described with the same words and behave very differently.
Questions
Can I teach Apple's dictation my colleague's name?
Not in the way that fixes the problem. macOS text replacement substitutes typed strings after the fact; it does not teach the recogniser how a name sounds. That is why names keep coming out wrong even after you add a replacement.
How many words should I add?
Fifteen to thirty. Adding hundreds makes accuracy worse, because rare words start winning against the common ones you actually say.
Does it work for non-English names?
Yes, and that is where it helps most. Give the spoken form the way you actually pronounce it — including an anglicised pronunciation if that is what you say — and the correct spelling with its accents.
Does the app learn automatically from my corrections?
In VV, no — entries are added deliberately. That is slower and far more predictable than software quietly inferring what you meant and being wrong about it.