Perceptually motivated blind source separation of convolutive audio mixtures with subspace filtering methods

Blind source separation aims to recover independent sources from their multiple observed mixtures using independent component analysis (ICA). However, when applying this technique to audio mixture problem such as a number of people talking in a room, the performance of the system is greatly reduced by the effect of room reflections and ambient noise. In contrast, two human microphones (the ears) perform well in such a real cock-tail party environment. Over the last few years the application of psycho-acoustic principles (the human auditory perception) has led to the successful development of MPEG audio coding standard which is the basic technique behind MP3 players and music availability on the internet. The first objective of this thesis is to apply psycho-acoustic principles to the spatial processing of speech signals in noisy and reverberant environment. The key assumption that will be adopted is that modern signal processing has failed to mimic the cock-tail party effect because there has been no attempt to adequately incorporate the psycho acoustical phenomenon of audio masking to aid source separation. A quasi linear mechanism for mimicking simultaneous frequency masking and temporal masking (post masking) techniques is developed. This framework is used to construct blind source separation algorithms that exploit audio masking prior to source separation (preprocessor) and after source separation (postprocessor). The final objective of this thesis is to exploit the perceptual irrelevancy of some of the input speech spectrum using the perceptual masking techniques before utilising the subspace method as a preprocessor of the frequency-domain ICA (FDICA) which reduces the effect of room reflections in advance and the remaining direct sounds then being separated by ICA. Incorporating the perceptual masking techniques prior to the application of MICA with the subspace method as preprocessor not only reduces the computational complexity of similarity measure for solving the permutations but also avoids the so-called permutation problem by targeting a specific speech signal more intelligible than the available microphone signals. Experiments carried out in both synthetic and real room scenarios and the results shown good objective performance in terms of signal-to-interference ratio (SW) and enhanced modified Bark spectral distortion (EMBSD) confirm the validity of the proposed solutions. Dedicated To My Parents