Emotion Recognition from Speech

Type: Master Thesis
Student: Rafael Schönenberger
Advisor: Cornelia Setz, Bert Arnrich
Project: SEAT

Research on emotions has a long history while automatic emotion detection represents a newer research field that is gaining more and more interest especially in Human Computer Interaction (HCI). In this thesis we specifically focused on acoustic emotion recognition from speech. Thus, we will give a theoretical basis on the features that were extracted from data and the classication applied to this data afterwards.

These features were processed by three different classification algorithms - linear discriminant analysis, decision trees and binary logistic regression, while binary logistic regression has not been used before in this field of classification. As speech data, recordings from a professional actors database was used in a first step. Additionally, an experiment with lay actors was done in the context of this thesis. Applying the database data as training data and the data obtained by the experiment as test data, respectively vice versa, the results for the different classification algorithms were computed. In a three-class problem the CART decision tree algorithm scored best with an accuracy of 50.4%, whereas the database data is used as training data and the lay actors data as test data.

Finally, a human recognition study reveals parallels and inequalities to the classification done by mathematical algorithms. In a six-class problem humans classified with an average accuracy of 53.02%. The classification algorithm that scored best in this context, the CART decision tree algorithm, classified with an accuracy of 32.1%.