Comparing Convolutional Neural Network and Support Vector Machine for Analyzing Anxiety Symptoms from Speech of Different Ethnicities

Abstract

Anxiety is a natural response of a person’s body to stress, but if a person experiences excessive anxiety on a regular basis, it could develop into a mental condition. In this study, two algorithms were utilized and compared in order to analyze anxiety symptoms from speech from a male gender. The audio database that was utilized for this study was the CREMA-D database and selected 1,152 male audio files with emotions that are related to anxiety, which were, Anger, Fear, Sad, and Disgusted. Audio pre-processing techniques were done to the audio files in order to improve the audio quality, as well as feature extraction techniques in order to obtain better accuracy from the models. The algorithms used in this study for recognizing anxiety symptoms were the SVM model and the Convolutional Neural Network. The SVM and CNN models performed well on the dataset, with accuracy scores of 62 percent and 78.1 percent, respectively, but it can be concluded that the CNN model outperformed the SVM model. For CNN, we were able to transform the audio datasets into pictures of heat maps of each audio file. It was then separated into two folders in an 80:20 ratio. While for the SVM, the researchers used the audio files themselves. The researchers used Python and BandLab as tools for the research.