ABSTRACT
Standard adaptive noise canceling uses linear filters to minimize the mean-squared difference between the filter output and the desired signal. For non-Gaussian signals, however, nonlinear filters can further reduce the mean-squared difference, thereby improving signal-to-noise ratio at the noise canceler output. This work investigates a two-microphone beamformer for suppressing directional background noise—an important task in, for example, radar, seismic or hearing aid applications. The beamformer includes an adaptive noise canceler with a nonlinear filter. Two nonlinear filters are examined: the Volterra filter (a specific sigma-pi neuron) and the multilayer perceptron. In the case of a single noise source emitting an independent, identically distributed (IID) random process, optimum linear and nonlinear performance limits are known for uniformly distributed noise. These limits were compared to the actual performance of the two nonlinear filters adapted off-line. The third-order Volterra filter and the perceptron with 20 hidden neurons performed equally well. For on-line adaptation, convergence speed and steady-state performance were scrutinized. In these experiments, the RLS-adapted Volterra filter outperformed the perceptron adapted with on-line backpropagation.
