Comparison of Different Talking Heads in Non-Interactive Settings

Six different talking heads have been evaluated in two consecutive experiments. Two text-to-speech components and three head components have been used. Results from semantic differentials show a clear preference for the most human-like and expressive head. The analysis of the semantic differentials reveals three factors each. These factors show different patterns for the head components. Overall quality is strongly related to one factor, which covers the quality aspect ‘appearance’. Another factor found in both experiments comprises ‘human likeliness’ and ‘naturalness’ and is much less correlated with overall quality. While subjects have been able to clearly separate all head components with different factors of the semantic differential, only some of these factors are relevant for explicit quality ratings. A good appearance seems to affect the perception of sympathy and the ascription of reliability.