A geometrically constrained multimodal time domain approach for convolutive blind source separation

Abstract

A novel time domain constrained multimodal approach for convolutive blind source separation is presented which incorporates geometrical 3-D coordinates of both the speakers and the microphones. The semi-blind separation is performed in time domain and the constraints are incorporated through an alternative least squares optimization. Orthogonal source model and gradient based optimization concepts have been used to construct and estimate the model parameters which fits the convolutive mixture signals. Moreover, the majorization concept has been used to incorporate the geometrical information for estimating the mixing channels for different time lags. The separation results show a considerable improvement over time domain convolutive blind source separation systems. Having diagonal or quasi diagonal covariance matrices for different source segments and also having independent profiles for different sources (which implies nonstationarity of the sources) are the requirements for our method. We evaluated the method using synthetically mixed real signals. The results show high capability of the method for separating speech signals.