Time-domain blind source separation for convolutive mixtures using second-order statistics

Abstract

Blind Source Separation (BSS) methods are generally used to separate the speech of people in the same room by using audio data recorded by multi-microphone system. This thesis focuses on a time-domain technique in BSS. In particular, the method presented uses the statistical independence of sources, the non-stationarity and the nonwhiteness of speech signals. The statistical independence of sources implies that crosscorrelations of demixed signals over all time-lags are null. Here, non-whiteness property is exploited by simultaneous minimization of output cross-correlations over multiple time-lags and non-stationarity property is exploited by simultaneous minimization of output cross-correlations at different time-instants. In the method, the demixing filter coefficients are obtained using an iterative optimization method and cost function. The coefficients are updated via the Natural Gradient adaptation method and the method is implemented in a block-online manner. Using Sylvester structure of the demixing filter matrix and with some approximations, the method is implemented efficiently. The method is found to be successful in simulations that are done using synthetically mixed signals in the computer. Moreover, the method is used to separate the noise from the speech signals that are also mixed synthetically and this suggests that the method can be used as a front-end processor for speech recognition applications.