Teager Mel and PLP Fusion Feature Based Speech Emotion Recognition

Although a number of features derived from linear speech production theory have been investigated as speech emotion indicators, the recognition accuracy still stays unsatisfactory for realistic applications. In this paper, Teager Mel, a novel speech emotion feature is proposed based on Teager Energy Operator (TEO) and the Mel perception characteristics. Due to such advantages as nonlinear and simple, TEO appears to be appropriate for speech emotion description. From the auditory psychophysical point of view, Perceptual Linear Predictive (PLP) features are also investigated as an extension to Teager Mel. A Support Vector Machine (SVM) classifier is then adopted to the fusion of Teager Mel and PLP features on a Chinese discrete emotional speech corpus (Dis-EC) that includes four emotions: happiness, anger, sorrow and surprise. Comparing with the previous studies based on prosodic features, the application of Teager Mel features can achieve a recognition accuracy improvement of 10.4%, and similarly 8.2% for PLP features. The recognition accuracy reaches79.7% while using the fusion features, which appears to be the most attractive in relative researches.