Artificial neural networks for nonlinear time-domain filtering of speech

Abstract

A multilayer perceptron (MLP) is applied as a time domain nonlinear filter to two classes of degraded speech, namely gaussian white noise and nonlinear system degradation introduced by a low bit-rate CELP coder. The goal of the study is to examine the influence of the inherent nonlinearity within the MLP, and this is achieved by varying the levels of nonlinearity within the structure. Direct comparisons of MLPs and linear filters show that with CELP degradation the SNR improvements achieved by the MLP is measurably better than with an equivalent linear structure (3 dB cf 1.5 dB) but when the degradation is additive noise the two structures perform equally well. The study highlights the importance of scaling to achieve optimum performance, and of matching the enhancer to the degradation.