Dynamic Minority Amplification and Fuzzy Fusion for Real-Time Stress Detection from Facial and Vocal Cues. 1Krishna Vamsi Pallapu, 2Dr. T. Santhi Sri

Abstract

The study is A real-time multimodal stress detection framework: The study fills the gap between affective computing and deployable mental health monitoring. The growing number of stress-related disorders and the lack of non-invasive, real-time detection systems stimulated the creation of the lightweight, but interpretable system that will identify stress using everyday facial and vocal expressions. Two benchmark datasets have been used: Affect Net reformatted to the YOLOv8-based facial emotion recognition, as well as CREMA-D, used to classify audio stress with a Multi-Task CNN-BiLSTM network with attention and residual connections. Both modalities were synchronized and integrated in time by a decision system based on fuzzy logic to make the interpretation and resistant to noisy or missing conditions of inputs. The presented fusion method reached the highest possible accuracy and F1-score outperforming unimodal baselines and offering natural and real-time interaction through producing a “Take a break” signal once continuous stress is detected. In general, the model proves the practicability of human-focused, emotional-focused, and cognitively intelligent AI in providing ongoing mental health care.