Multi-Stage Approach for Stress Detection Using Speech Lexical Analysis

Stress is one of the psychological problems that people most often experience. Stress that is not managed well and occurs for a prolonged time can lead to depression, negatively impacting a person's physical and mental health. Detecting and identifying stress is challenging, but it is necessary to provide appropriate early treatment for someone who experiences it. Many researchers have explored various technology-based approaches to detecting stress based on these needs. One approach currently being actively researched is stress detection technology using voice signals as an indicator to detect stress in a person. However, of these various studies, studies have yet to focus on exploring the use of voice properties optimally in terms of its properties as a signal wave and its output as a linguistic product. In this paper, we propose a multi-stage stress detection model through speech using a combination of speech signal characteristics and lexical aspects of speech. We employ the Mel-Spectrogram for the sound feature extraction required for stress detection from speech signals. Then, using Word2Vec and TF-IDF as input representations, we developed an NLP model for stress detection via lexical aspects of speech. The evaluation results of our proposed multi-stage stress detection model were able to detect stress well by showing an average accuracy and F1-Score of 91.7 and 91.2, respectively.