Article ID: 2026EAP1041
To enhance feature extraction in speech signals and prevent critical scale information from being diluted by redundant noise, we propose an AdaBoosted Multi-scale & Multi-feature Model (AMSF). The AMSF model integrates multi-scale convolution with AdaBoost and introduces a global perception fusion module, which helps to emphasize informative emotional features while suppressing noise. The model extracts MFCC and Logfbank features from speech signals and uses Pearson's correlation coefficient for feature fusion, enhancing discriminative power. In addition, DiffGrad optimization is adopted to stabilize the training process, accelerating convergence and improving model performance. Experimental results on the improvised subset of the IEMOCAP dataset show that the AMSF model achieves Weighted Accuracy (WA) of 86.17%, Unweighted Accuracy (UA) of 83.03%, Macro F1 score of 83.69%, and Weighted F1 score of 85.80%.