A Feature Fusion Model with Data Augmentation for Speech Emotion Recognition

The Speech Emotion Recognition (SER) algorithm, which aims to analyze the expressed emotion from a speech, has always been an important topic in speech acoustic tasks. In recent years, the application of deep-learning methods has made great progress in SER. However, the small scale of the emotional speech dataset and the lack of effective emotional feature representation still limit the development of research. In this paper, a novel SER method, combining data augmentation, feature selection and feature fusion, is proposed. First, aiming at the problem that there are inadequate samples in the speech emotion dataset and the number of samples in each category is unbalanced, a speech data augmentation method, Mix-wav, is proposed which is applied to the audio of the same emotion category. Then, on the one hand, a Multi-Head Attention mechanism-based Convolutional Recurrent Neural Network (MHA-CRNN) model is proposed to further extract the spectrum vector from the Log-Mel spectrum. On the other hand, Light Gradient Boosting Machine (LightGBM) is used for feature set selection and feature dimensionality reduction in four emotion global feature sets, and more effective emotion statistical features are extracted for feature fusion with the previously extracted spectrum vector. Experiments are carried out on the public dataset Interactive Emotional Dyadic Motion Capture (IEMOCAP) and Chinese Hierarchical Speech Emotion Dataset of Broadcasting (CHSE-DB). The experiments show that the proposed method achieves 66.44% and 93.47% of the unweighted average test accuracy, respectively. Our research shows that the global feature set after feature selection can supplement the features extracted by a single deep-learning model through feature fusion to achieve better classification accuracy.