|Supervisor:||Dr.-Ing. Robert Aichner|
|Faculty:||Prof. Dr.-Ing. Walter Kellermann|
Blind source separation (BSS) addresses the problem to separate sources from a set of linear mixtures. In the acoustical scenario this corresponds to multiple speech signals acquired by a microphone array. Thus the mixing system is convolutive, i.e. it consists of room impulse responses which can be modelled by finite impulse response (FIR) filters. To accomplish separation we have to assume that the source signals are mutually independent, which is in general true for speech signals. The term "blind" stresses that no additional a priori knowledge about the source signals, mixing conditions or the sensor array configuration is necessary.
In most standard BSS tasks a quadratic case is assumed, i.e., the number of sensors is equal to the number of sources. Recently, many researchers have been dealing with the more difficult scenario that there are more source signals than sensors. This scenario is especially important when dealing with speech signals in real-world environments. Many different approaches have been proposed, mainly exploiting the so-called sparseness of the source signals in a suitable transformation domain. Sparseness means that regions in the transformation domain can be determined where only one source signal is active. Additionally another class of algorithms is trying to solve this problem by extracting the sources of interest whereas the rest of the sources is neglected. This technique is called "blind signal extraction".
In this master thesis promising approaches for the underdetermined BSS problem should be investigated. First experiments should verify the algorithms in the free-field case, i.e., the mixing filters consist only of a delay and attenuation. Further work will include the examination of the algorithms in reverberant enclosures.