Integration of DFT and cosine-modulated filter banks with blind separation of convolutively mixed non-stationary sources

In this paper, oversampled M̃ channel FIR filter banks using both DFT modulation and cosine modulation designs are used in conjunction with a time domain blind source separation (BSS) algorithm [1]. This BSS algorithm has been shown to blindly separate the fullband versions of non-stationary convolutively mixed sources in the time domain. However further savings on convergence and computational complexity can be made by using subband decomposition on the mixed signals before implementation of the time domain BSS algorithm in each subband. An extended lapped transform (ELT) prototype is modulated using a cosine-modulated (CM) FTR filter bank and then with a DFT modulated FIR filter bank. Both of these designs are compared to the typical frequency domain BSS approach to solving these convolutive non-stationary BSS problems such as in [2]. The signal to interference ratio (SIR) is used as the performance metric to evaluate and analyse the comparison of the three separation methods.