Historically the acoustic models have tended to range from around 12mb to 16mb and the current English model shipping with the distribution is 16mb. It doesn’t get recompiled with separate slices, so it will increase the size of the app by the exact size of the model. There isn’t really anything in there that can be reduced IMO.
Additionally, is it possible to only limit the size of the model to a couple of words in order to shrink the model?
Well, the acoustic model itself doesn’t contain words (other than the phonetic dictionary I’ll get into more below), but it is used in combination with a language model which is where the vocabulary words are specified. OpenEars is designed to allow you to very easily create and switch between the smallest-necessary language models (or rules-based grammars) so that language model size is always slim and the vocabulary only consists of words relevant to the user’s current task, so it isn’t necessary to take any special steps in order to accomplish that since it’s the default behavior.
The acoustic model does also ship with a master phonetic dictionary which is a word list for finding pronunciations for new language models, and that isn’t a small file, but it is part of what makes the dynamic language model generation create accurate models, so I think there isn’t a lot of advantage in reducing it – you might save 1mb or so but you’d tend to generate less-accurate models as a result.