Neural network filters for speech enhancement

In adaptive noise cancelling, linear digital filters have been used to minimize the mean squared difference between filter outputs and the desired signal. However, for non-Gaussian probability density functions of the involved signals, nonlinear filters can further reduce the mean squared difference, thereby improving the signal-to-noise ratio at the system output. This is illustrated with a two-microphone beamformer for cancelling directional interference. In the case of a single uniformly distributed interference, we establish the optimum nonlinear performance limit. To approximate optimum performance, we realize two nonlinear filter architectures, the Volterra filter and the multilayer perceptron. The Volterra filter is also examined for speech interference. The beamformer is adapted to minimize the mean squared difference, but performance is measured with the intelligibility weighted gain. This criterion requires the signal-to-noise ratio at the beamformer output. For the nonlinear processor, this can only be determined when no target components exist in the reference channel of the noise canceller so that the target is transmitted without distortion. Under these ideal conditions and at equal filter lengths, the quadratic Volterra filter improves the intelligibility-weighted gain by maximally 2 dB relative to the linear filter