It is possible to evaluate how good an N-gram model is without integrating it into an automatic speech recognition. We simply measure how well it predicts some unseen test data.
in Dan Jurafsky and James H. Martin “Speech and language processing: an introduction to natural language processing, computational linguistics, and speech recognition”, 2009, Pearson Prentice Hall, Upper Saddle River, N.J., Second edition, ISBN 0135041961
This is the new version. Still under construction.