Stress is a natural physiological and psychological response to demanding situations, but accurate detection remains challenging in real-world scenarios. Existing stress databases are limited and often recorded under extreme conditions or acted, which makes real-world applications difficult. This study introduces a multi-modal stress database, naturally elicited through student evaluations, incorporating normal and thermal video alongside speech recordings annotated by psychologists. While physiological signals provide robust stress detection, speech and facial expressions are more accessible. We extract physiological parameters from speech to bridge this gap by automatically detecting breath locations. These breath indices are then used to derive speech breathing parameters. Machine learning techniques classify these parameters for stress detection. Our proposed method marginally outperforms conventional speech-based features, demonstrating a promising approach to stress analysis.