Cascaded emotion classification via psychological emotion dimensions using a large set of voice quality parameters

In this paper we improve the speaker independent emotion classification of a well known German database consisting of 6 basic emotions: sadness, boredom, neutral, anxiety, happiness, and anger. We achieve this by adding a large set of voice quality parameters to the standard prosodic features. In addition, our observation that the optimal feature set strongly depends on the emotions to be classified, leads to a 3-stage cascaded classification motivated by the psychological model of emotion dimensions. After activation recognition in the first stage, we classify the potency and evaluation dimension in the second and third stage, respectively. Compared to the 2-stage approach, the average classification rate is improved by 14% to 88.8%.