Speech emotion recognition using supervised manifold learning based on all-class and pairwise-class feature extraction

Supervised manifold learning has been used widely in speech emotion recognition, but dimensionality is reduced with all classes being considered jointly, difference between feature subsets of different classes is ignored. In this paper, new idea about feature extraction based on fusion of all-class and pairwise-class is proposed, an improved manifold learning method is used to extract features and the feature subsets is divided into several parts, one is based on all-class structure, others based on each pair of classes. All-class subset is used in K-nearest (KNN) classifiers, others used in Support Vector Machine (SVM) classifiers, and an supervised multi-classifiers system of speech emotion recognition is constructed. Experiments show a significant improvement in recognition accuracy.