Identification of Psychological Stress from Speech Signal Using Deep Learning Algorithm

Psychological stress has become a major issue in modern society due to increasing academic, social, and professional pressures, and early detection is critical to prevent serious mental health consequences. Existing systems predominantly rely on traditional assessment approaches such as surveys, questionnaires, and expert psychological evaluations, which are time-consuming, inherently subjective, and require trained clinical professionals. To overcome these limitations, this paper proposes an automated stress detection system built upon BERT (Bidirectional Encoder Representations from Transformers), a state-of-the-art deep learning model that analyzes usergenerated text to classify whether content indicates stress or non-stress conditions. The system processes text through preprocessing, tokenization, and fine-tuned model training using Natural Language Processing (NLP) techniques. The technology stack includes Python, BERT, Deep Learning, NLP, PyTorch, the Transformers library, and Streamlit for a user-friendly web interface enabling realtime text analysis and probability visualization. Through optimized training strategies including stratified data splitting, learning rate scheduling, dropout regularization, and balanced datasets, the system achieves an accuracy exceeding 95%, making it a reliable and scalable tool for early psychological stress detection in educational, workplace, and healthcare environments.