Simulating Human-to-Human Conversations for Prediction of Conversational Quality
* Presenting author
Degradations in current telephone transmissions like packet-loss or delay have a multitude of repercussions on a conversation. In real world scenarios instrumental estimations of transmission with delay and packet-loss do not take into account the type of conversation that is being measured. In this paper, we propose a new approach to predict the perceived quality of transmissions affected by delay and packet-loss by simulating the impacts these degradations have on a conversation. For this, we use user-simulation techniques from the spoken dialogue community to model human conversational and turn-taking behaviour in two types of scenarios, namely Short Conversation Tests and Random Number Verification Tests. We show how the impact of packet-loss may be modelled by simulating the misunderstanding of information, and the respective restructuring of the conversation to resolve that misunderstanding. We also propose a way intended and unintended interruptions can be simulated with turn-taking rules, and how the simulated agents might resolve the conflicts that arise from these interruptions. We then outline possible quality measures that may be derived from such a system, including a mean opinion score for degraded conversations.