关键词: 关键词: 语音情感识别, 渐进式情感表征学习, 卷积神经网络, 动态融合, 多尺度融合

Abstract: Speech emotion recognition aims to equip computers with the ability of accurately identifying emotional states conveyed in speech signals, and how to efficiently represent emotional information has been a key research focus. Currently, most research efforts are dedicated to leveraging deep learning methods to directly learn optimal features from raw speech or spectrogram data. These approaches are good at extracting more comprehensive representation information. However, they may overlook the learning of more refined information within specific features and may not guarantee feature interpretability. Therefore, CnnPRL, a convolutional neural networks-based progressive representation learning method, is proposed for speech emotion recognition. Firstly, interpretable shallow features are hand-crafted and the optimal feature set is selected. Secondly, a cascaded convolution network and dynamic fusion structure are proposed to refine the shallow features and learn deep emotional representations. Lastly, a parallel heterogeneous CNN network is constructed to extract complementary information at different scales by utilizing a fusion module to achieve multi-feature fusion, to capture multi-granularity features, and integrate deep emotional information from different feature scales. Under the premise of ensuring time complexity, the experimental results on the datasets IEMOCAP, CASIA, and EMODB demonstrate that, compared to BiGRU-Focal，TLFMRF and TIM-Net, CnnPRL achieves 1.63%, 2.92%, and 2.82% improvement in WAR index, respectively, and it is proven to be effective. Ablation experiments demonstrate that each module of the CnnPRL contributes positively to the overall performance of the model.

Key words: Keywords: speech emotion recognition, progressive emotional representation learning, convolutional neural network, dynamic fusion, multi-scale fusion