An optimal model with a lower bound of recall for imbalanced speech emotion recognition

In an early complain warning system, we encounter a common problem - the lack of angry emotions for training classification models. Moreover, the recognition of angry emotion is more important than that of no-anger emotion. Based on this, the main purpose of this paper is to train an optimal model which achieves a high recall above a lower bound and a maximum of F₁ score. It is divided into three aspects: 1) A variant of F₁ score (TF₁ score) takes recall above a lower bound and F₁ score into consideration; 2) A Single Emotion Deep Neural Network (SEDNN) and its training process are designed to find an optimal model with a maximum of TF₁ score. 3) A performance comparison of different methods is conducted on IEMOCAP and Emo-DB database. Extensive experiments show that when a BCE loss function or a focal loss function is used, the training process can find a model with a recall above a high threshold and a maximum of F₁ score. Especially, SEDNN with the focal loss function performs better than SEDNN with the BCE loss function.