F-SEONS: a second-order frequency domain algorithm for convolutive source separation in noisy environments [speech recognition applications]

Convolutive source separation is a problem of restoring original unknown sources from multiple sensor signals that are modelled as convolutive mixtures of mutually independent sources. It is a fundamental problem which plays a critical role in cocktail party speech recognition. Most of existing methods neglected the effect of sensor noise in the model of convolved mixtures. In this paper, we consider the case of noisy data and present a method of source separation, F-SEONS, which jointly exploits the nonstationarity and temporal structure of sources in the frequency-domain. Numerical experiments, comparing to other methods, confirm the high performance of our proposed method.