Large-scale Language Models for Conversational Speech Recognition