Speech Emotion Recognition (SER) has emerged as a key sector in the sphere of human-computer interaction, aiming to extract emotional cues from speech signals. Identifying effective emotion features for speech signals remains a challenge due to the scarcity of data. This paper presents an innovative approach to address these issues. The proposed method first calculates Mel-frequency Spectral Coefficients (MFSC) features and their derivatives, and then segments them into equal-length segments using a dynamic step-size method. This segmentation strategy ensures the retention of original speech information. A feature fusion model, inspired by the gate structure in LSTM, selectively focuses on emotion-related segments. Additionally, a segment noise-label suppressed module is introduced to refine segment labels. Experiments on IEMOCAP dataset show that the proposed model attains an accuracy of 70.37%, outperforming several cutting-edge algorithms.