A supervised non-negative matrix factorization model for speech emotion recognition

Abstract

Feature representation plays a critical role in speech emotion recognition (SER). As a method of data dimensionality reduction, Non-negative Matrix Factorization (NMF) can obtain the low-dimensional representation of data by matrix decomposition, and make the data more distinguishable. In order to improve the recognition ability of NMF for SER, we conduct a potential study on NMF and propose a supervised NMF model, called joint discrimination ability and similarity constraint of NMF (DSNMF). This model incorporates the discriminative information and similarity information of samples into basic NMF as prior knowledge, so that the original data can be decomposed into more distinguished low-dimensional data. Specifically, on the one hand, the labels of the training set are used to improve the discriminative ability of the model; on the other hand, with the similarity of the training samples, the data of similar samples are more highly aggregated in the low-dimensional space. In addition, the convergence of DSNMF is proved theoretically and experimentally. Extensive experiments on EMODB and IEMOCAP corpora show that the proposed approach has a better classification effect on low-dimensional representation data than other NMF models.