Emotional expression analysis in vocal performance is of great significance for enhancing artistic expression, and the rapid development of artificial intelligence technology provides a new solution path for music emotion recognition. This paper proposes a strategy for analyzing and enhancing emotional expression in vocal performance based on pattern recognition, and constructs a multimodal music emotion classification model based on optimized residual network. Methodologically, Mel spectrum is used for audio signal preprocessing, GloVe word vectors are utilized to represent the lyrics text, the model performance is enhanced by teacher-student modeling and transfer learning, the ResNet50 network structure is improved and an improved Center-Softmax classifier is introduced. Experimental results on the classical piano dataset show that the proposed algorithm achieves 88.34% emotion recognition accuracy, which is a 2.43% improvement compared to the XGBoost algorithm, and the recall and F1 values reach 83.34% and 86.52%, respectively. In the Chinese folk song multi-emotion recognition experiment, the recognition accuracy of the multimodal fusion model reaches 85.62%, which is 6.15% and 4.19% higher than the unimodal model, respectively. The vocal performance visualization analysis verifies the effectiveness of the model in ancient poetic art songs such as Guan Ju. The experiments proved that the method can effectively recognize multiple emotional states in vocal performance, providing scientific technical support for vocal teaching and performance evaluation.
