Paul Shields, Mark Girolami, Douglas Campbell, Colin Fyfe, Dept of Electronic Engineering and Physics, Dept of Computing and Information Systems, University of Paisley, High Street, Paisley PA1 2BE, Scotland
This paper reports on two engineering techniques for the enhancement of speech corrupted with noise and reverberation. Features of the human auditory system, in particular the binaural unmasking effect inspires both techniques. Motivated by the primary processing of the cochlea a diverse sub-band adaptive processing (DSBAP) technique employing signal cancellation is described. The use of frequency limited sub-bands within the processing mechanism can be compared with the critical frequency bands found within the mammalian cochlea. The second technique is additionally inspired by the binaural unmasking effect, which appears to utilise binaural correlation properties to de-emphasise an undesired signal. Motivated by temporally sensitive hebbian super-synapses, which may form sparse representations based on commonly occuring spatio-temporal input patterns, we propose a form of temporal anti-hebbian learning within a laterally connected network. This is applied, as a wide-band method to the enhancement of noise corrupted speech signals and is empirically compared to the sub-band processing scheme.