On the Role of Conversational Timing in Synthetic Training Data for ASR