Cross-corpus Speech Emotion Recognition Using Transfer Semi-supervised Discriminant Analysis

Many speech emotion recognition approaches have been presented in recent years, and most of them assume that emotional speech utterances in training and testing corpora are collected under the same conditions. However, in many real applications, this assumption does not hold as the training data and testing data are often obtained from different scenarios, e.g., ages, noises, languages. To address this problem, in this paper, a novel transfer learning approach, called transfer semi-supervised linear discriminant analysis (TSDA), is presented for cross-corpus speech emotion recognition. On one hand, the distribution similarity between source and target databases is considered. On the other hand, the semi-supervised linear discriminant analysis (SDA) algorithm is adopted for feature dimension reduction. Finally, the transfer SDA method, which jointly optimizes the SDA and distribution similarity measurement together, is proposed. Experiments are carried out on public emotional datasets, and results demonstrate the effectiveness of our proposed approach.