Real world source separation by combining ICA and VD-CDWT in time-frequency domain

Abstract

It is well known that in real world source separation, the environment noise removal must be considered with complex reverberating sound, and various noises. In this study, in order to improve the voice recognition accuracy in real world source separation, a new method that uses Independent Component Analysis (ICA) in the time-frequency domain using the variable density complex discrete wavelet transform (VD-CDWT) and the subspace method has been proposed. Through comparison of the results according to signal noise ratio (SNR), the effectiveness of the proposed method is confirmed.