Aichner R (2007)
Publication Language: English
Publication Type: Thesis
Publication year: 2007
City/Town: University Erlangen-Nuremberg, Germany
In the last few years blind source separation techniques have received a great deal of attention in the signal processing community. One reason is that blind source separation approaches only rely on the assumption of mutual statistical independence of the source signals and do not need additional a-priori information, such as the sensor array geometry or the positions of the desired and interfering sources. Moreover, several separated point sources are retrieved simultaneously instead of extracting one desired source as common in beamforming. In this work we focus on acoustic signals and present several important special cases of a generic blind source separation framework which was termed TRINICON ("TRIple-N Independent component analysis for CONvolutive mixtures"). This framework is based on an information-theoretic criterion allowing to incorporate also higher-order statistics into the adaptation algorithms, which is in contrast to beamforming methods solely relying on second-order statistics. Moreover, this framework allows a unified view on convolutive blind source separation algorithms leading to novel algorithms and showing also relationships to popular state-of-the-art algorithms. After explaining the framework, we present some approximations which lead to highly efficient algorithms while still preserving the superior properties of the general framework relative to other known algorithms. A second major contribution is that, beyond existing blind source separation literature, we address the application of blind source separation to reverberant and noisy environments by presenting several pre- and post-processing techniques in a coherent treatment. The algorithms allow to maintain a high separation performance of the blind source separation algorithms also in noisy scenarios and additionally, are capable of suppressing the undesired diffuse background noise which cannot be treated by the convolutive blind source separation model. Experimental results demonstrate the applicability of the proposed approaches in real-world environments such as reverberant rooms, which may also exhibit background noise, and in the particularly significant case of noisy passenger compartments.
Aichner, R. (2007). Acoustic Blind Source Separation in Reverberant and Noisy Environments, PhD thesis (Dissertation).
Aichner, Robert. Acoustic Blind Source Separation in Reverberant and Noisy Environments, PhD thesis. Dissertation, University Erlangen-Nuremberg, Germany, 2007.