Edinburgh Napier University

  This paper presents non-linear adaptive speech enhancement schemes inspired by features of early auditory processing. A generic multi-microphone sub-band adaptive (MMSBA) framework is described which allows for the manipulation of several factors that may influence the intelligibility and perceived quality of the processed speech. The proposed framework supports inclusion of: non-linear distribution of sub-bands (as in humans), cross-band effects such as lateral inhibition, and robust adaptive metrics for selecting an appropriate coherent or incoherent noise canceller for each sub-band, based on identified features of the band-limited signals from multiple-sensors during silence periods. An efficient higher order statistics (HOS) based speech/non-speech detector is proposed for enabling effective adaptive control of MMSBA filtering against the environment. New hybrid extensions of the MMSBA scheme incorporating neural networks and post-Weiner filtering are also described and their comparative performance assessed in real reverberant environments. Finally, some future research directions for MMSBA based speech enhancement are proposed including possible alternative strategies based on stochastic resonance.