New Hybrid Sub - band Speech Enhancement Systems incorporating Neural Networks and post - Weiner Filtering

In this paper, two new hybrid sub-band systems are proposed which aim to combine neural network sub-band processing with post-Wiener filtering for adaptive speech-enhancement processing of noisy signals. The proposed hybrid architectures comprise an early auditory-processing modelling inspired Multi-Microphone Sub-band Adaptive (MMSBA) system incorporating neural-network based non-linear sub-band filters, integrated with post-Wiener filtering (WF) in order to further reduce the residual incoherent noise components resulting from the application of conventional non-linear MMSBA processing (without WF). A human cochlear model resulting in a nonlinear distribution of the sub-band filters (as in humans) is also employed in the developed schemes. Preliminary comparative results achieved in simulation experiments using anechoic speech corrupted with real automobile noise show that the proposed structures are capable of significantly outperforming the conventional non-linear MMSBA and wide-band noise cancellation schemes.