Speech Emotion Recognition based on DCNN BiGRU Self-attention Model

Abstract

In speech emotion recognition, the features extracted by handmade design are generally low-level, they may not be enough to distinguish subjective emotions, and speech signals are usually have time sequence and every frame signal has a different role. Therefore, this paper aims at the above problems, a DCNN BiGRU self-attention model is proposed. The model combines the spatial characteristics of convolutional neural networks, the advantages of circulating neural network in learning time series data, and the characteristics of attention mechanisms that can learn feature weights, thereby improving the accuracy of speech emotion recognition. This model achieved an average recognition rate of 89.53% and 91.74% in the EMO-DB and CASIA databases, and through comparison with other literatures, it is proved that this model can obtain more ideal results in speech emotion recognition.