Speech-Based Stress Classification based on Modulation Spectral Features and Convolutional Neural Networks

Interest in stress recognition has notably increased over the past few years. In this work, we focus on recognizing stress from speech. We propose the use of modulation spectral features as input to a convolutional neural network (CNN) for classifying stress. As benchmark, the OpenSMILE features used in the INTERSPEECH 2010 Paralinguistic Challenge is adopted and evaluated with a support vector machine (SVM) and a deep neural network (DNN) based backends. Experiments are performed with the well-known Speech Under Simulated and Actual Stress (SUSAS) database. Performances are investigated considering 2-class, 4-class and 9-class classification problems. Results show that the proposed approach outperforms the benchmark on a challenging 9-class classification task with accuracy as high as 70% representing gains of roughly 18% over the benchmark.