Unravelling stress levels in continuous speech through optimal feature selection and deep learning

Abstract

Stress is a psychological or emotional strain that occurs due to adverse experiences in human life. This paper showcases the application of deep learning in detecting stress levels in continuous audio signals in the Distress Analysis Interview Corpus Wizard of Oz (DAIC-WOZ) database. The features that have been experimented with are Gammatone Frequency Cepstral Coefficients (GFCC), Log Filter Bank (Log-Filter Bank), Mel Frequency Cepstral Coefficients (MFCC), chroma, and Linear Predictive Coding (LPC). Five deep learning models were evaluated: Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), Bidirectional LSTM (Bi-LSTM), k-fold CNN with the k value as 5, and a fusion model of CNN, LSTM, and attention. Upon evaluating the performance metrics of all the models, it is concluded that the k-fold CNN model with k as 5 performs well with continuous audio signals. The model has achieved an accuracy of 80% when it is trained on MFCC, GFCC, and Log-F Bank features which are observed to be the optimal features in the stress analysis.