Improvements in Blind Separation of Speech Signals in Reverberant Rooms

Blindly separating multiple signals from mixed speech sources recorded within a reverberant room is a difficult task, which still requires improvement to algorithms proposed in the literature to accelerate convergence and reduce interference, distortions and artifacts present in the outputs. This paper addresses this problem by considering a determined mixing system, in which the numbers of sources and sensors are equal, and exploiting non-stationarity, non-Gaussianity and non-whiteness of speech signals, an approach known as Trinicon (Triple-N Independent Component Analysis). By utilizing ambient impulse response (RIR) attributes, which can be estimated blindly through machine learning techniques, the proposed method improves the solution of the multi-parameter optimization problem by incorporating prior knowledge of the characteristics of an optimal solution. Simulation results are presented, comparing the performance of the proposed approach with that of the conventional method using simulated impulse responses in a room with varying reverberation times and speech signals.