Neural Networks and TEO Features for an Automatic Recognition of Stress in Spontaneous Speech

This study presents automatic stress recognition methods based on acoustic speech analysis. Novel approaches to feature extraction based on the nonlinear Teager energy operator (TEO) calculated within critical bands, discrete wavelet transform bands, and wavelet packet bands are presented. The classification process was performed using two types of neural networks: the multilayer perceptron neural network (MLPNN) and the probabilistic neural network (PNN). The classification efficiency was tested using the actual stress dataset from the SUSAS database. The speech recordings were made by 15 speakers (8 females and 7 males) reading a list of 35 words under three actual conditions: high stress, low stress, and neutral. The best overall performance was observed for the features extracted using the TEO parameters calculated within perceptual wavelet packet bands (TEO-PWP). Depending on the type of mother wavelet, the correct classification scores for the PWP features ranged from 71.24% to 91.56% (using the MLPNN classifier), and from 86.63% to 93.67% (using the PNN). The PNN classifier outperformed the MLPNN classification method.