Speech based analysis of physiological stress using perceptually significant features

Analysis of physiological stress based on features from speech parameters is preferable over other techniques because of its nonintrusive nature. In this work, heart rate (HR) at three different levels (low, medium and high heart rates) is used as a measure of stress and a common speech utterance in each case is employed to correlate stress with HR. Using a database consisting of utterances and their corresponding heart rates and other physiological measurements indicative of stress for the speakers, utterances of the word eye for a speaker at low, medium and high heart rates are isolated. From these short utterances, features based on auditorily significant areas are extracted for analysis. Spectral features that are perceptually significant, i.e., those above global masking threshold in each band, appear to distinguish speech at three different heart rates at a better rate than the commonly used cepstral features or the total energy in each Bark frequency band. Between the two pattern matching techniques used, namely, dynamic time warping and neural network, the latter showed a higher accuracy in classifying the heart rates from all the features employed. Based on the small set of training and test utterances, the neural network based feature classification technique demonstrated the capability of the proposed set of speech features in classifying physiological stress.