@InProceedings{Kroschel2007_471,
author = {Kristian Kroschel and Michael Grimm and Vasilije Krstanovic},
booktitle = {Studientexte zur Sprachkommunikation: Elektronische Sprachsignalverarbeitung 2007},
title = {Evaluation of Vocal-Facial Based Emotion Primitives},
year = {2007},
editor = {Klaus Fellbaum},
month = mar,
pages = {35--42},
publisher = {TUDpress, Dresden},
abstract = {Emotion detection gains improved importance in man-machine in-
teraction. There are two approaches to represent emotion: first, discrete ele-
mentary emotions like fear or joy might be defined, or an emotion space is used
in which the discrete emotions define a point or a cluster. In this paper the
latter method is used with a three-dimensional space spanned by the emotion
primitives valence, dominance, and activation.
Synchronously, sequences of speech and facial expressions are extracted from
a TV talk-show to generate an audio-video data base of authentic emotions.
From these recordings, three data subsets have been generated: video alone,
audio alone and combined audio-video. Investigations have shown that the
emotions expressed by facial parameters change with a time constant of roughly
1 sec. Thus an additional subset of segmented video data has been generated.
These data have been evaluated by 15 persons, i.e., they estimated the inten-
sity of the emotion primitives in the range from -1 to +1 using so-called Self
Assessment Manikins (SAM). To test the reliability of these estimates, the
correlation coeffcients of the estimates of the individual evaluators have been
calculated. As a second measure, the standard deviation of the estimates for
the four subsets and the three emotion primitives has been calculated. These
measures have been used to decide whether or not the evaluated data base can
be used as the ground truth for automatic emotion classification.},
isbn = {978-3-940046-40-6},
issn = {0940-6832},
keywords = {Grundlagen},
url = {https://www.essv.de/pdf/2007_35_42.pdf},
}