Synthesizing Attitudes in German

Our study investigates the potential of modeling the vocal expression of human attitudes based on a limited set of prosodic and voice quality parameters and their subsequent synthetic realization. Four attitudes (uncertainty, sincerity, surprise and doubt) were taken into account. For two utterances, a set of acoustic prosodic (F0, intensity, duration) and voice quality parameters (jitter, shimmer) were extracted from a corpus of German expressive speech and analyzed acoustically, yielding a set of voice adaptation rules for attitudinal modeling. The subsequent synthesis and paralinguistic voice adaptation was carried out using MaryTTS HMM voices embedded in the InproTK system. In a first objective evaluation, a comparison of lexically identical human and synthesized expressive utterances yields mostly positive correlations between the acoustic parameters used for analysis and modeling. To find out whether these similarities are sufficient to create the perceptual impression of the examined attitudes, a subjective evaluation was carried out: Listeners were asked to identify a target attitude in pairs of synthesized utterances either characterized by a rising, interrogative contour (doubt vs. surprise) or a falling, declarative contour (sincerity vs. uncertainty). That way, uncertainty was identified in 90%, followed by sincerity (80%), surprise (72%) and doubt (64%).