The redundant information, noise data generated in the process of single-modal feature extraction, and traditional learning algorithms are difficult to obtain ideal recognition performance. A multi-modal fusion emotion recognition method for speech expressions based on deep learning is proposed. Firstly, the corresponding feature extraction methods are set up for different single modalities. Among them, the voice uses the convolutional neural network-long and short term memory (CNN-LSTM) network, and the facial expression in the video uses the Inception-Res Net-v2 network to extract the feature data. Then, long and short term memory (LSTM) is used to capture the correlation between different modalities and within the modalities. After the feature selection process of the chi-square test, the single modalities are spliced to obtain a unified fusion feature. Finally, the fusion data features output by LSTM are used as the input of the classifier LIBSVM to realize the final emotion recognition. The experimental results show that the recognition accuracy of the proposed method on the MOSI and MELD datasets are 87.56 and 90.06%, respectively, which are better than other comparison methods. It has laid a certain theoretical foundation for the application of multimodal fusion in emotion recognition.