Abstract. This paper focuses on the selection and justification of deep learning models for emotion classification tasks. It provides a comprehensive analysis of various neural network architectures, including Convolutional Neural Networks, Recurrent Neural Networks, Long Short-Term Memory networks, and Transformer models, assessing their performance in recognizing and classifying human emotions from multimodal data sources. The study examines the strengths and limitations of each model with respect to data type, training efficiency, computational complexity, and generalization capabilities. Furthermore, criteria for optimal model selection tailored to real-world emotion recognition applications are discussed. The findings contribute to enhancing the accuracy and robustness of emotion classification systems and offer valuable guidelines for researchers and practitioners developing advanced affective computing solutions.