Fixes for accuracy issues with language models containing punctuation, and cases in which stuck utterances are incorrectly calculated.