Speech Processing for Hearing Aids: Noise Reduction Motivated by Models of Binaural Interaction

Several signal processing techniques are reviewed that aim at reducing ambient noise and enhancing the "desired" speech signal in complex acoustical environments ("cocktail-party processing"). These algorithms are motivated by models of binaural interaction in the normal human auditory system and try to simulate several different aspects of normal auditory function that are typically impaired in hearing-impaired listeners. All algorithms assume input signals from microphones located near the ears of a subject and one or two output signals to be presented. The first class of algorithms performs a directional filtering with respect to the forward direction and a reduction of the perceived reverberation. The second class of algorithms performs an analysis in the modulation frequency domain and combines binaural cues with cues from modulation frequency analysis to perform a noise-robust directional filtering. The third class of algorithms simulates a localization process in a way comparable to neurophysiological findings in the barn owl, while the fourth class of algorithms combines cues from binaural interaction and fundamental frequency analysis. The respective psychoacoustical and physiological motivation of these algorithms as well as their advantages and shortcomings are outlined. In addition, the hardware and software required for implementing and testing these algorithms in real-time are introduced and discussed. Since most of these algorithms are shown to provide significant benefit by increasing the "effective" signal-to-noise ratio in different acoustical situations, a combination of these algorithms appears promising for future "intelligent" digital hearing aids.