Speech Emotion Recognition Based on Transfer Emotion-Discriminative Features Subspace Learning

Cross-corpus speech emotion recognition(SER) is a hot topic in classification. SER includes these four issues:feature selection, differences constraint, label regression and preservation of discriminative features. Seldom literature can solve issues jointly previous studies.In this work,we propose the transfer emotion-discriminative features subspace learning(TEDFSL) method.Acoustic are extracted by OpenSMILE source target data. Then sent into CNN+BLSTM to learn higher-level global time series. The common low-dimensional data learned Linear Discriminant analysis (LDA) reduce dimension Maximum Mean Discrepancy (MMD) Graph Embedding (GE) constraint between low- dimensional combined with matrix relationship labels features,after which the, DNN selected as final classifier preserve features, emotion-aware center loss( $\mathrm {l}_{\mathrm {c}}$ ) added extensive experiments carried out on cross-corpus tasks results demonstrate that our proposed method superior state-of-art SER.