Various NMF analyses for emotion recognition

To extract emotion from speech signals, we must specify the representation and grammatical model, which are still challenging issues. We proposed a new feature called Nonnegative Matrix Factorization (NMF) feature. The proposed algorithm has been tested in several different ways by varying NMF and the speech database. We compared the recognition rate only Euclidian distance with enhancing ways (SVM, Partial multiplication of vectors, SFM) of the NMF classification. Observing all these together, we found that total recognition rate is improved. We concluded that the NMF feature indicates both spectral information and temporal information, which is an efficient tool over other spectrum-based features.