Neural network learning of spectral features of nonverbal speech

Abstract

A neural-network approach to the learning of invariant spectral features in cerebral-palsied speech is introduced. The technique is a hybrid conventional digital signal processing/neural network strategy. The objective at this stage is to learn features of nonverbal speech, and to do so in a manner which is robust to the abnormalities of such speech and which is minimally dependent on a priori modeling or parameterization. Thus, it is hoped that these features will be useful as input to a higher-level word recognizer.<>