Comparing Features for Acoustic Anger Classification in German and English IVR Portals

Acoustic anger detection in voice portals can help to enhance human computer interaction. In this paper we report about the performance of selected acoustic features for anger classification. We evaluate the performance of the features on both a German and an American English dialogue voice portal database which contain “real” speech, i.e. non-acted, continuous speech of narrow-band quality. Deploying a large-scale feature extraction we determine the optimal set of features for each language. To obtain the ranking we use an Information-Gain Ratio filter. Analyzing the most promising features we notice a predominance of MFCC and loudness features. However, for the English database also pitch features proved importance. We further calculate classification scores for our setups using discriminative training and Support-Vector Machine classification. The developed systems show that Emotion Recognition in both English and German language can be processed very similarily.