Extracting Emotional Feature Changes in College Music Performances via Multi-Channel Convolutional Attention Mechanism

To address the limitations of traditional emotional-feature extraction methods for musical performances, including low accuracy, inadequate temporal-continuity modeling, and weak differentiation of subtle emotional changes, this study proposes a multi-channel convolutional attention-based method for extracting emotional feature variations. A multimodal data acquisition system for university music performances is first constructed to collect audio data, including pitch, rhythm, and dynamics, and visual data, including facial expressions and body movements, from vocal and instrumental performances. A multi-channel convolutional neural network is then designed for multidimensional feature extraction from heterogeneous data. The audio channel uses 1D-CNN to capture time-domain and frequencydomain features, while the visual channel combines 2D-CNN and 3D-CNN to extract spatial and temporal sequence features. A hybrid attention mechanism integrating channel attention and temporal attention is introduced to adaptively weight key emotional features and enhance the capture of dynamic emotional changes. Experiments on the self-built UMPED dataset and the public MusicNet dataset evaluate emotion-feature similarity, classification accuracy, and temporal emotion-change fitting, verifying the effectiveness of the proposed multimodal signal-processing framework.