On the Modelling of Prosodic Cues in Synthetic Speech – What are the Effects on Perceived Uncertainty and Naturalness?

In this paper we present work on the modelling of uncertainty by means of prosodic cues in an articulatory speech synthesizer. Our stimuli are embedded into short dialogues in question-answering situations in a human-machine scenario. The answers of the robot vary with respect to the intended level of (un)certainty, the independent variables are intonation (rising vs. falling) and filler (absent vs. present). We perform a perception study in order to test the relative impact of the prosodic cues of uncertainty on the perception of uncertainty and also of naturalness. Our data indicate that the cues of uncertainty are additive. If both prosodic cues of uncertainty are present, the perceived level of uncertainty is higher as opposed to the deactivation of a single cue. Regarding the relative contribution of intonation vs. filler our results do not show a significant difference between judgments. Moreover, the correlation between the judgment of uncertainty and of naturalness is not significant.