| Résumé |
In this chapter, a review of theories that deal with controlness of emotional reactions raises the difficulty of evaluating the degree of spontaneity in an emotion by observing only its external expression. Although empirical methods involving direct and indirect measurements have been used to collect emotional data, it is difficult for them to identify the inherent degree of control the speaker has on his emotional reaction, whether it is simulated or not. Therefore emotions and their corresponding expressions have to be distinguished. We provide an attempt at defining expressivity as an information level of communication. It involves all perceivable demonstrations, controlled or not, of an internal state that can be interpreted as uncontrolled. Expressivity can formally be separated from other information levels, e.g. speaker identity, speaking style, modality and prominence. In practise, the task is harder since all these information levels share the same paralinguistic channel composed of restructurings of verbal words and non verbal sounds, all modulated by prosody. This problem can be solved by discarding prosodic variations due to other information levels, e.g. involving the so-called neutral case. By the use of contextual clustering an expressive French corpus, significant prosodic variations can be attributed to the change of the expressivity. A statistical context-dependent method is described. The resulting speaker independent model of expressivity is used to modify the expressivity of a neutral utterance, either recorded or synthesized. The proposed method is aimed at artistic ends, but also participates to the anthropomorphisation of Human-machine interfaces. |
