| Blind source separation (BSS) attempts to recover a set of statistically independent sources from a set of mixtures knowing only the structure of the mixing network, and the hypothesized probability distribution function of the sources. The case where the sources are immobile persons speaking in a reverberant room is of particular interest, because it represents a first step toward unlocking the so-called "cocktail party problem". Due to the reverberations, BSS in the time domain is usually expensive in terms of computations, but the number of computations can be significantly decreased if separation is carried out in subbands.; An implementation of a subband-based BSS system using DFT filter banks is described, and an adaptive algorithm tailored for subband separation is developed. Aliasing present in the filter bank (due to the non-ideal frequency response of the filters) is reduced by using an oversampled scheme. Experiments, conducted with two-input two-output BSS systems, using both subband and fullband adaptation, indicate that separation and distortion rates are similar for both systems. However, the proposed 32-subband system is approximately 10 times computationally faster than the fullband system. |
