The research is dedicated to developing an information system that utilizes multimodal data (audio, visual, and textual) for the recognition and analysis of a person's emotional and mental state. The relevance of this research stems from the constant stress faced by many individuals in challenging environments, affecting their psychological well-being. The main goal is to create a system that enables users to receive psychological assessments and support through a digital interface, avoiding the discomfort some people experience when interacting with specialists. The system integrates facial expression analysis, speech intonation, and text processing to provide comprehensive emotion recognition. The research applies a multimodal neural network approach using a large and diverse dataset containing over 65 h of unscripted video content. Various neural network architectures, such as CNN, LSTM, and transformer-based models, are compared to determine the most effective structure for emotion recognition. The system processes input data using specialized encoders, then combines them in a generative model for predicting emotions and mood. The study involves using multimodal data and a large, real-world dataset. Practical applications include real-time emotion recognition through a client–server interface, tested on real users via an interactive interface. The results demonstrate that the multimodal approach significantly improves accuracy compared to unimodal models, showing the effectiveness of combining visual, audio, and textual data for emotion prediction. The system achieved an accuracy of 51% in real, non-studio conditions, indicating its reliability. This highlights the potential for further development into a fully functional tool for psychological assistance and mental state assessment.