Innovation in the field of human–computer interaction involves analyzing users’ real-time emotions, which stands to be an essential and challenging task as they can be easily controlled or faked. Methodologies for analyzing emotions in existing studies include facial, audio, and physiological signals. The primary objective is to develop a model for emotion classification that can accurately identify and interpret human emotions through skin temperature, respiration, and plethysmograph. The aim was to analyze ensemble models that accurately discern and interpret emotional states. The emotional states were classified based on the frequency domain signal components extracted using Fast Fourier Transform (FFT), such as amplitude and frequency. Ensemble-based machine learning algorithms such as XGBoost and LGBM achieved the highest accuracy in classifying various emotional states. The study involves unimodal and bimodal analysis of the signals. The comparative classification rate of bimodal results is the highest for calm, with 85.5%, by combining a plethysmograph and temperature. Whereas the bimodal results with respiration and skin temperature maintain the accuracy level for all four emotions. The results also convey the significance of plethysmograph and temperature for a high classification rate of happiness and emotion, whereas respiration has improved the classification rate of anger and sadness. The potential applications include enhancing user experiences and contributing valuable insights into mental health care, human–computer interaction, and recommendation systems.