Considering Idling Motion Generation using RNNs for Virtual Conversational Agents

Abstract

This work aims to develop a model to generate fine grained and reactive non-verbal behaviors of the virtual character when the human user is talking to it. The target non-verbal idling behaviors are micro facial expression, head movements, and postures. We explored the use of recurrent neural network (RNN) to learn these behaviors in reacting to the human communication interlocutor's corresponding micro non-verbal behaviors. The models are trained on an active listening data corpus which features elderly speakers talking with young active listeners and was collected by ourselves.