What a vocabulary pack is and is not

Twenty one vocabulary packs ship with Zumbo, covering fields like developer tools, legal, medical, design, marketing, writing and research, and video and audio. People reasonably assume this means a list of words somewhere. It is worth explaining what a pack actually does, because the difference is the reason it works.

What it is not

It is not a find and replace applied to the output. macOS already has that, under Keyboard, as text replacement, and it operates after the words have been recognised. That is useful for a term you can predict, and useless for the general case, because by the time the text exists the mistake has already been made and the original sound is gone.

It is also not a dictionary you have to fill in. There is no screen in Zumbo that asks you to type your profession’s terminology before you start, and there is no cap on how many words a field is allowed to have.

What it is

A pack biases the recogniser towards the vocabulary of a field while it is deciding what it heard. A general model is choosing, at every moment, between candidate words weighted by how likely they are in ordinary language. In ordinary language, a library name is not a word at all, so it loses to the everyday phrase that sounds like it. The pack changes those weights for the terms your field actually uses.

The practical consequence is that the correction happens before the text exists, rather than being patched afterwards. That is why a pack fixes cases a replacement rule cannot: it is not correcting the output, it is changing the decision.

Why there are twenty one and not one

Because the weights are the whole point, and they conflict. A word that should be likely in a clinical note should not be likely in a commit message. Turning on every field at once would push the model towards a soup of jargon from professions you do not work in, which is measurably worse than leaving it alone.

So Zumbo asks during setup what you dictate most, switches on the matching packs, and leaves the rest off. You can change them at any time in Settings, under Dictionary.

Where packs stop and Teach starts

Packs handle the shared language of a field, which is the part that is the same for everybody in it. They cannot help with the words that exist only inside your company: internal service names, customer names, the codename for the project, your colleagues’ surnames. Nothing trained on public text has ever seen those.

That is what Teach is for. Correct the word once, in place, and it is heard your way from then on, with the correction stored on your Mac. Packs and Teach stack: the first gives you the profession, the second gives you the job.

More on that in how Teach corrections work, and the full set is on the vocabulary packs page.

Devtools 173 terms Marketing 329 terms Medical 89 terms Legal 87 terms Design 95 terms Finance 106 terms + 15 more