Co-speech Gesture Generation for Embodied Agents and its Effects on User Evaluation - Coverbal Synchrony in Human-Machine Interaction - page 235

a questionnaire in which two types of dependent variables had to
be assessed. First, participants were asked for their evaluation of
the presentation quality with respect to Max's verbal eloquence,
gestures' helpfulness, the comprehensibility of Max's explanations,
and the vividness of the agent's mental image of the church. Second,
participants were asked to state their person perception of the embodied
agent in terms of items like 'polite', 'authentic', or 'cooperative'. In
the analysis, these items were grouped into scales for warmth and
competence.
It turned out that both individual GNetIc conditions outperformed
the other conditions in that gestures were perceived as more helpful,
overall comprehension was rated higher, and the agent's mental
image was judged as more vivid (see Figure 4). Similarly, the two
individual GNetIc networks outperformed the control conditions
regarding agent perception in terms of warmth and competence. The
combined GNetIc condition was rated worse than the individual
GNetIc conditions in all of these categories. Further, the no gesture
condition was rated more positively than the random condition, in
particular for the subjective measures of overall comprehension, the
gesture's role for comprehension, and vividness of the agent's mental
image. That is, with regard to these aspects it seems better to make
no gestures than to randomly generate gestural behavior even though
it is still reasonably iconic.
Overall, individualized gesturing was strikingly beneficial with
regard to how embodied agents and their communicative skills are
judged by human users, suggesting that modeling individual speakers
with proper abilities for the target behavior results in even better
1.5
3.0
4.5
6.0
Figure 4. Participants' ratings of the embodied agent Max regarding the quality of the
produced iconic gestures and the agent's personality.