New Variants of Nonnegative Matrix Factorization with Application to Speech Coding and Speech Enhancement
Sammanfattning: In this thesis, new variants of nonnegative matrix factorization (NMF) based ona convolutional data model, -divergence and sparsication are developed andanalyzed. These NMF variants are collectively referred to as -CNMF. Commonsparsication techniques such as L1-norm minimization and elastic net arediscussed and a new regularizer is proposed. It is shown that the new regularizer,unlike the above-mentioned sparsication techniques, has control overthe number of active bases in the NMF dictionary. Moreover, the -CNMF isextended to multichannel signals: it learns a common dictionary by exploitingthe correlation between channels through a multichannel coecient matrix. Asa result, an algorithm for source separation based on multichannel -CNMF isdeveloped. The algorithm is further tested in a multilayer setting, in which thefrequency-shifted coecient matrices serve as input to the next higher layer.Finally, three variants of the algorithm are evaluated in the context of speechenhancement, focusing on the problem of speech extraction from complex auditoryscenes. Figures obtained from the SiSEC 2016 data show that the proposedalgorithms perform comparably or better than the state of the art.
HÄR KAN DU HÄMTA UPPSATSEN I FULLTEXT. (följ länken till nästa sida)