Exploiting non-negative matrix factorization with linear constraints in noise-robust speaker identification

Abstract

This paper exploits non-negative matrix factorization (NMF)-based method for speech enhancement within speaker identification framework. The proposed algorithm considers speech atoms in deterministic way as a sum of harmonically-related sinusoids in spectral domain. This approach allows us to estimate specific signal structure of vowel signal in the presence of noise in order to make an efficient noise reduction using only noise exemplars. The experiments of the present research in application to the speaker identification are conducted on the computational hearing in multisource environments (CHiME) dataset. The obtained results demonstrate the effectiveness of the preprocessing enhancement, and outperforming the general NMF-based speech enhancer. Further studies show the channel compensation effect of the proposed method leads to performance comparable to the common mismatch reduction methods such as feature warping.