Aiming at the problems of low recognition accuracy and poor generalization ability of current multimodal emotion recognition algorithms in modal feature extraction and information fusion between modalities, a multimodal emotion recognition algorithm based on speech, text and expression is proposed. Firstly, a shallow feature extraction network (Sfen) combined with parallel convolution module (Pconv) is designed to extract the emotional features in speech and text. A modified Inception-ResnetV2 model is adopted to capture the emotional features of expression in video stream. Secondly, in order to strengthen the correlation among modalities, a cross attention module is designed to optimize the fusion between speech and text modalities. Finally, a bidirectional long and short-term memory module based on attention mechanism (BiLSTM-Attention) is used to focus on key information and maintain the temporal correlation between modalities. By comparing the different combinations of the three modalities, it is found that the hierarchical fusion strategy that processes speech and text in advance can obviously improve the accuracy of the model. Experimental results on the public emotion datasets CH-SIMS and CMU-MOSI show that the proposed model achieves higher recognition accuracy than the baseline model, with three-class and two-class accuracy reaching 97.82% and 98.18% respectively, which proves the effectiveness of the model.