Representative and Discriminant Feature Extraction Based on NMF for Emotion Recognition in Speech

For the emotion recognition in speech we had developed two feature extraction algorithms, which emphasize the subtle emotional differences while de-emphasizing the dominant linguistic components. The starting point is to extract 200 statistical features based on intensity and pitch time series, which are considered as the superset of necessary emotional features. Then, the first algorithm, rNMF (representative Non-negative Matrix Factorization), selects simple features best representing the complex NMF-based features. It first extracts a large set of complex almost-mutually-independent features by unsupervised learning and latter selects a small number of simple features for the classification tasks. The second algorithm, dNMF (discriminant NMF), extracts only the discriminate features by adding Fisher criterion as an additional constraint on the cost function of the standard NMF algorithm. Both algorithms demonstrate much better recognition rates even with only 20 features for the popular Berlin database.