- Hendrik Barfuß
In many applications for speech and audio signal enhancement schemes, the desired signal is corrupted by some amount of unpredictable interference and noise. If several sensors are available, signal enhancement is typically based on multichannel noise reduction schemes exploiting second-order statistics. A prominent example for second-order statistics-based multichannel signal enhancement is the Linearly-Constrained Minimum Variance (LCMV) filtering scheme, which allows for linear spatial constraints up to the number of available sensors to suppress interfering sources and to preserve desired source(s). For an efficient realization of LCMV filters, the Generalized Sidelobe Canceller (GSC) is commonly employed, which turns the constrained optimization problem into an unconstrained problem.
For nonstationary wideband signals such as speech, adaptive versions of the GSC are necessary, which, however, require a robust control mechanism to adapt both the blocking matrix and the interference canceller in order to minimize signal distortion of the extracted signal(s). This well-known sensitivity of the GSC to signal cancellation results from reflection paths of the desired source(s).
In this thesis, a novel realization of the GSC based on geometrically constrained Independent Component Analysis (ICA) should be analyzed and evaluated theoretically as well as by Matlab simulations. Geometrically constrained ICA is exploited to adaptively estimate both the constraint and the blocking matrix under reverberant conditions when only an imprecise knowledge of the desired source position is available. For different multisource conditions, both the convergence behavior and the obtained result of ICA with and without constraints should be analyzed. Moreover, the influence of the weight of the geometric constraint on ICA needs to be evaluated as well as the robustness against errors in the required target source position estimate. This novel realization of the GSC should also be compared to recent GSC realizations [Talmon et al. 2009]. Well-documented and well-structured software is important.