Training a model that will be used only in grammar mode

When training a model for other language that will be used only in grammar mode, do I need less data compared to models being trained for dictation? If so, can you estimate how many hours of audio do I need?

In my use case the words that will be used in the grammar and not known to me and will be entered by the user of the app. Also, the grammar should include "unk" for unrecognized words.

Thanks