In the speech mode， the OpenSMILE toolbox is used to extract low-level acoustic features from the speech signal. Transformer Encoder is richer to excavate deep features from low level acoustic features and fuses them so as to obtain more useful emotional representation. In the text mode， considering the association between pause and emotion， the speech and text are aligned to obtain the pause information and the pause information is added to the transcript text by pause encoding. The utterance-level lexical features are obtained by the improved DC-BERT model. Then，acoustic features and lexical features are fused and the bi-directional long short-term memory based on attention neural network （BiLSTM-ATT） is used for emotion classification. Finally， this paper compares the effects of three different attention mechanisms integrated into BiLSTM on emotion recognition （local attention， self-attention and multi-headed attention），and local attention is found to be the most effective. In the experiments on IEMOCAP dataset， the method proposed in this paper achieves 78.7% in weighted accuracy for four emotion categories， which is better than the baseline system.