Speech emotion recognition is a crucial technology in human-computer interaction. However, it faces such challenges as insufficient feature selection, limited temporal modeling capability, and high complexity of emotional expression. To address these issues, a novel model called CoAttention-guided Temporal Capsule Network (CATC-Net) is proposed. It integrates a collaborative attention mechanism, bidirectional scalable long short-term memory network (Bi-sLSTM), and capsule networks for comprehensive modeling of multi-dimensional emotional features. The model begins with a feature desensitization mapping block to enhance generalization in complex and noisy environments. Next, a channel-wise weighting strategy based on multi-head attention is introduced to jointly model and fuse global contextual information with channel-level importance. The Bi-sLSTM structure further improves the capture of speech rhythm and dynamic emotional changes. Finally, the capsule network compresses and encodes high-level semantic features while preserving their compositional relationships. Extensive experiments on datasets IEMOCAP, EMODB, CASIA, and BodEMODB show that the proposed CATC-Net achieves higher accuracy than existing methods, demonstrating strong robustness and effective emotion recognition.