ESSV Konferenz Elektronische Sprachsignalverarbeitung

Title: Speaker-Group Specific Acoustic Differences in Consecutive Stages of Spoken Interaction

Authors: Ronald Böck, Olga Egorow, Andreas Wendemuth

Abstract:

The communication of humans is influenced by various circumstances, especially in real-life situations. A particular aspect is the detection of affective states in an interaction and their changes. An automatic detection of such changes is challenging and highly related to significant features used for the detection. Therefore, we analysed the differences in spectral and prosodic features in an interaction, in particular, in two distinct communication stages. This investigation was done on a subset of 18 participants of the LAST MINUTE Corpus providing several different affect afflicted stages in a naturalistic Human-Computer Interaction. Across the 18 participants, 16 of the 52 analysed features – a subset of the emobase feature set – showed significant differences between the two considered stages. Most remarkably, seven of them are related to Mel-Frequency Cepstral coefficients. Furthermore, we found that there is a subset of five participants showing similar feature changes in the two stages. Four of them belong to the elder speaker-group, regardless of their sex. The presented findings confirm that features showing similar changes across and within speaker-groups as well as speaker-group dependent modelling are advantageous in the assessment of affective Human-Computer Interaction.


Year: 2017
In session: Dialogsysteme
Pages: 211 to 218