Wavelet packet energy and entropy features for classification of stressed speech

This paper presents an approach that aims to recognize stress under speech. The proposed system is based on wavelet packet prosody features. These features are extracted from speech according to Mel scale, Bark scale and ERB scale. Multiclass Support Vector Machines are used as the based classifiers in order to classify the stress states. The speech utterances used is this study are taken from Speech Under Simulated and Acted Stress database. The classification accuracies obtained when we apply wavelet energy and entropy features exceed 98% when we use the Gaussian kernel with the multiclass SVM approaches.