F-SEONS: a second-order frequency-domain algorithm for noisy convolutive source separation

We present a frequency-domain method of noisy convolutive source separation, where we extend the SEONS algorithm (Choi et al. (2002)) that jointly exploits the nonstationarity and temporal structure of sources. Thus, the method is called F-SEONS, implying frequency-domain SEONS. Unlike most of the existing methods of convolutive source separation, we consider the case of noisy data and show that our F-SEONS algorithm identifies multivariate FIR channels in a robust way. In addition, we employ an H_∞ filtering method in order to further suppress the sensor noise. Numerical experiments compared to other methods confirm the high performance of our proposed method.