Speech enhancement using non-linear prediction

Describes a method for speech enhancement using multi-layer perceptrons (MLP) configured for time domain non-linear prediction (NLP) as originally proposed by Kaouri and Lin (1988) and Tamura and Waibel (1988). In addition, the authors present details of the influence of the non-linear operations compared directly with the linear case. Experimental results show that nonlinear MLPs provide better performances than linear equivalent systems. Results also show that MLPs with a multi-output structure give improved performances over MLPs with single-output structure and normalisation can affect MLPs performance. >