In the field of speech emotion recognition, the Time Perceived Bidirectional Multiscale Network (TIM-NET) faces two main challenges: on the one hand, existing models fail to fully capture the spatial details in speech data, which limits the in-depth extraction of emotional information from speech signals; on the other hand, the lack of diversity and complexity in emotional speech data makes it difficult for the model to discern subtle differences between various emotional expressions. To address these issues, this paper proposes an improved model framework that integrates time-series processing techniques with spatial feature extraction methods to comprehensively capture the spatiotemporal correlations in speech data. Specifically, this study extends the TIM-NET module by incorporating Graph Convolutional Networks (GCNs) to extract spatiotemporal features from speech data, thereby enhancing the model’s ability to understand speech signals. Additionally, this paper introduces the Wasserstein Generative Adversarial Network (WGAN) to generate diverse and complex emotional speech data, which helps the model better learn and distinguish subtle differences between different emotions. Experimental results demonstrate that the proposed model outperforms the TIM-NET in terms of speech emotion recognition performance on the RAVDESS, EMO-DB, and EMOVO datasets, showcasing stronger emotion recognition capabilities and better generalization performance.