Anger or Joy? Emotion Recognition Using Nonlinear Dynamics of Speech

Abstract

Recent developments in man–machine interaction have increased the need for recognizing human emotion from speech. The present study aimed to classify the highly confused emotions, anger and joy, using Nonlinear Dynamic features (NLDs). The proposed NLDs are extracted from the geometrical properties of reconstructed phase space of speech. A linear support vector machine is employed to classify emotional speech signals. The recognition rates of 99.1% and 98.85% were achieved on the Berlin database for females and males, respectively. The proposed system can also be employed as an error-correction procedure to reduce ambiguity between anger and joy in multiemotional problems. We show that applying this strategy on a selected multiemotional system significantly improves the overall recognition rate from 91.59% to 94.58%. These results reveal the capability of the proposed NLDs to classify highly confused emotions, joy and anger.