When InkSpoke keeps mangling a product name or an acronym, there are three different mechanisms that could fix it. They behave differently, and picking the wrong one is why people conclude "vocabulary does not work".

The three systems

System	Scope	Can it rewrite text?
Custom vocabulary	Global, syncs across devices	Yes — with a replacement
Workspace vocabulary	One workspace	No — biasing only
Personal dictionary	Global, local to the device	Yes — including snippet expansion

Custom vocabulary

This is the one to reach for first. Entries come in two forms:

example · one per line

kubernetes
cubernetties => Kubernetes
ink spoke => InkSpoke

A bare term biases transcription toward that spelling. A term => replacement pair also rewrites the transcript afterwards, catching the cases where the speech model got it wrong anyway. Replacement is whole-word and case-insensitive, so it will not corrupt words that merely contain the term.

Custom vocabulary sits behind the global layer master switch on the Personal Context page, which is off by default. If your vocabulary appears to do nothing at all, check that switch before anything else.

Workspace vocabulary — and the inheritance trap

Each workspace has its own term list, and a setting controlling how it combines with the global one. That setting has three values:

Mode	Effect
Inherit (default)	Uses the global list only — this workspace's own terms are ignored
Append	Workspace terms on top of the global list
Workspace only	Only this workspace's terms

This is the single most common vocabulary complaint. Typing terms into a workspace while it is left on Inherit does nothing — the list is stored but never used. Switch it to Append for those terms to take effect.

Workspace vocabulary is bias-only. It cannot rewrite text, so replacement pairs must go in the global custom vocabulary.

Personal dictionary

The dictionary is a trigger-and-replacement list applied after transcription, and it also handles snippets — say a short trigger, get a longer block of text. Longer triggers are matched first, and your personal entries take precedence over any team entries.

The budget nobody mentions

Bias terms are passed to the speech model as a prompt with a hard ceiling of 1,024 characters. Terms beyond that are silently dropped, in a defined order:

Workspace terms go in first
Then global vocabulary
Then dictionary triggers — these are dropped first when space runs out

✺

The settings page shows a live character count against that 1,024 limit and warns at 80%. If you are near it, prune generic words — bias only helps for terms the model would otherwise get wrong, so "meeting" is wasted budget and "Kubernetes" is not.

Separately, at most 50 terms are passed to the refinement model as known terms.

A note on secrets

Vocabulary and personal context are scanned for values that look like credentials. If one is detected, that data is refused for cloud sync and the app tells you rather than uploading it quietly. Vocabulary is not a password manager — do not put keys in it.