Surface electromyography (sEMG) is a crucial type of bioelectrical signal, captured by sensors placed on the surface of skin, which reflects the activity state of muscles. sEMG-based gesture recognition involves identifying muscle action patterns from these signals, with potential applications in fields like biomechanics, rehabilitation, prosthetic control, and human-computer interaction. However, previous studies typically feed raw time-series signals directly into the network. It remains unclear whether incorporating time-frequency domain signals can enhance feature extraction, which motivates the research presented in this paper. To address this issue, we propose a deep learning architecture for time-frequency domain sEMG gesture recognition that integrates a convolutional neural network (CNN), a bidirectional gated recurrent unit (BiGRU), and a transformer module. Specifically, we obtain time-frequency domain signals using the short time fourier transform (STFT), which are then fed into a Transformer-based CNN-BiGRU network for feature classification. Unlike previous work that uses a cascaded structure of CNN and GRU, we adopt a parallelized structure of CNN and BiGRU, followed by a Transformer classifier. By utilizing such a parallel structure, spatial and temporal features are extracted separately and then fused before being sent to the Transformer, achieving effective temporal-spatial feature fusion. Experimental results show that the proposed architecture outperforms the benchmark model in the gesture recognition task.