Speech Emotion Recognition Based on 1D-CNNs-LSTM Hybrid Model

Speech emotion recognition is a key technology for achieving human-computer interaction. Implementing efficient emotion classification has always been a research topic for researchers. In response to the problem of weak representation of local feature information extracted from the original speech segments, and considering the complexity of manually extracting features, in order to learn the best feature representation from the original speech segments and extract contextual information from cumbersome speech segments, a 1D convolutional neural networks long short-term memory (ID-CNNs-LSTM) neural network model has been proposed. The two Local feature enhancement block (LFEB) modules in the new model further enhance local feature information, improve data expression ability and classification performance. The LSTM module can effectively extract contextual information from speech segments, which helps to improve the accuracy of the model. The new model can automatically extract features from original speech segments and has the advantages of small parameter size, easy training, and low computational complexity, makes it more suitable for low-power and low memory devices. The new model was tested on the Berlin EMODB dataset and achieved an optimal classification accuracy of 94.49%, which is 2.47% higher than the 2D CNN model, verifying the effectiveness of the model.