Deep Convolutional Neural Network and Gray Wolf Optimization Algorithm for Speech Emotion Recognition

Speech emotion recognition (SER), an important method of emotional human-machine interaction, has been the focus of much research in recent years. Motivated by powerful Deep Convolutional Neural Networks (DCNNs) to learn features and the landmark success of these networks in the field of image classification, the present study aimed to prepare a pre-trained DCNN model for SER and provide compatible input to these networks by converting a speech signal into a 3D tensor. First, using a reconstructed phase space, speech samples are reconstructed in a 3D phase space. Studies have shown that the patterns formed in this space contain meaningful emotional features of the speaker. To provide an input that is compatible with DCNN, a new speech signal representation called Chaogram was introduced as the projection of these patterns, and three channels similar to RGB images were obtained. In the next step, image enhancement techniques were used to highlight the details of Chaogram images. Then, the Visual Geometry Group (VGG) DCNN pre-trained on the large ImageNet dataset is utilized to learn Chaogram high-level features and corresponding emotion classes. Finally, transfer learning is performed on the proposed model, and the presented model is fine-tuned on our datasets. To optimize the hyper-parameter arrangement of architecture-determined CNNs, an innovative DCNN-GWO (gray wolf optimization) is also presented. The results of this study on two public datasets of emotions, i.e., EMO-DB and eNTERFACE05, show the promising performance of the proposed model, which can greatly improve SER applications.