Classroom student emotion recognition using an improved segmentation clustering and multi-feature fusion emotion recognition algorithm

To address noise, speech masking, and weak robustness in classroom emotion recognition, this study proposes a model combining enhanced segmentation clustering and multi-feature fusion.An improved U-Net with local loss supervision first performs denoising.Secondly, using MFCC features combined with Bayesian segmentation and K-means clustering to process speech signals.Finally, MFCC, formant, and pitch features are integrated into an attention-based BiLSTM for emotion recognition.Results show the U-Net achieved a loss of 0.26 after 16 iterations, with PESQ at 3.01 and STOI 84.21%.Segmentation false negative and positive rates were 13.84% and 12.52%.K-means purity reached 90.85%.The multi-feature model attained 92.98% accuracy for excited emotion, and the full system reached 93.62% test accuracy.The model improves recognition in complex classrooms, supporting personalised smart education.