Stress detection using non-semantic speech representation

In today's world, stress has become a prominent cause for many ailments. Automatic detection of stress from speech using state-of-the-art machine learning algorithms can facilitate early detection and prevention of stress. Artificial intelligence agents involved in affective computing and human-machine spoken interaction (HMI) might benefit from the capacity to identify human stress automatically. Despite the fact that several different methods have been established for stress detection, it is still unclear which auditory features should be considered for training a deep neural network (DNN) model. In this study, we propose to investigate the performance of traditional and modern auditory features for stress classification using the StressDat database. The StressDat database is a collection of acted speech recordings in Slovak realizing sentences within stress-prone situations in three different levels of stress. The performance of traditional auditory features such as Mel-Frequency Cepstral Coefficients (MFCC) and Perceptual Linear Prediction (PLP) are compared with modern auditory non-semantic speech representation such as x-vectors and TRIpLet Loss network (TRILL) vectors. As a benchmark, Low-level descriptors (LLD) auditory features are extracted using the OpenSMILE toolkit. We evaluated performance of four different automatic classification algorithms: support vector machine (SVM), multilayer perceptron (MLP), convolutional neural network (CNN), and long shortterm memory (LSTM). The results reveal that TRILL vectors trained on CNN provide the highest accuracy (81.86%).