Speech Enhancement by Classification of Noisy Signals Decomposed Using NMF and Wiener Filtering

Supervised non-negative matrix factorization (NMF) is effective in speech enhancement through training spectral models of speech and noise signals. However, the enhancement quality reduces when the models are trained on data that is not highly relevant to a speech signal and a noise signal in a noisy observation. In this paper, we propose to train a classifier in order to overcome such poor characterization of the signals through the trained models. The main idea is to decompose the noisy observation into parts and the enhanced signal is reconstructed by combining the less-corrupted ones which are identified in the cepstral domain using the trained classifier. We apply unsupervised NMF followed by Wiener filtering for the decomposition, and use a support vector machine trained on the mel-frequency cepstral coefficients of the parts of training speech and noise signals for the classification. The results show the effectiveness of the proposed method compared with the supervised NMF.