Speech Emotion Recognition using XGBoost and CNN BLSTM with Attention

Speech emotion recognition is an essential part of enhancing human-computer interaction, and its recognition accuracy depends on the selection of speech signal features and the construction of classification models. In this paper, we propose a speech emotion recognition framework. It is based on the XGBoost algorithm for feature selection, and combines the ID convolutional neural networks and the BLSTM (bi- directional long short-term memory) with attention model for classification. By calculating the importance score of each feature, the appropriate number of effective features can be obtained to alleviate the over-fitting problem, reduce the training overhead, accelerate the modeling process and improve the recognition accuracy. The proposed framework is evaluated on EmoDB, CASIA, EMA corpora, and is shown to provide more accurate predictions compared to the baseline method by achieving the recognition accuracy of 86.87%, 74.17%, 98.04%, respectively.