Head Motion Generation

Abstract

Head movement is an important part of body language. Head motion plays a role in communicating lexical and syntactic information. It conveys emotional and personality traits. It plays an important role in acknowledging active listening. Given these communicative functions, it is important to synthesize Conversation Agents (CAs) with meaningful human-like head motion sequences, which are timely synchronized with speech. There are several studies that have focused on synthesizing head movements. Most studies can be categorized as rule-based or data-driven frameworks. On the one hand, rule-based methods define rules that map semantic labels or communicative goals to specific head motion sequences, which are appropriate for the underlying message (e.g., nodding for affirmation). However, the range of head motion sequences that are generated by these systems are usually limited, resulting in repetitive behaviors. On the other hand, data-driven methods rely on recorded head motion sequences which are used either to concatenate existing sequences creating new realizations of head movements or to build statistical frameworks that are able to synthesize novel realizations of head motion behaviors. Due to the strong correlation between head movements and speech prosody, these approaches usually rely on speech to drive the head movements. These methods can capture a broader range of movements displayed during human interaction. However, even when the generated head movements may be tightly synchronized with speech, they may not convey the underlying discourse function or intention in the message. The advantages of rule-based and data-driven methods have inspired several studies to create hybrid methods that overcome the aforementioned limitations. These studies have been proposed to generate the movements using parametric or nonparametric approaches, constraining the models not only on speech, but also on the semantic content. This chapter reviews most influential frameworks to generate head motion. It also discusses open challenges that can move this research area forward.