Importance-driven turn-taking for spoken dialogue systems

Turn-taking is a critical aspect to any spoken dialogue system as it governs the tim- ing and nature system utterances. Standard turn-taking approaches are overly rigid and unprincipled, only seeking to mimic the surface features of human turn-taking instead of addressing the underlying motivation behind the behavior. Here, we introduce Importance- Driven Turn-Taking which combines the importance of speaking with a variable strength turn-taking signal using reinforcement learning. We describe and evaluate theoretical, architectural, and practical aspects of the approach in both simulation and in live user experiments. Overall, we find that the importance-driven approach is more efficient than traditional methods as it can exhibit more flexible behaviors that can accommodate different types of users.