Effective Cross-Utterance Language Modeling for Conversational Speech Recognition