Assessment of a Speaker Recognition System Based on an Auditory Model and Neural Nets

This paper deals with a new speaker recognition system based on a model of the human auditory system. Our model is based on a human nonlinear cochlear filter-bank and Neural Nets. The efficiency of this system has been tested using a number of Spanish words from the 'Ahumada' database as uttered by a native male speaker. These words were fed into the cochlea model and their corresponding outputs were processed with an envelope component extractor, yielding five parameters that convey different auditory sensations (loudness, roughness and virtual tones). Because this process generates large data sets, the use of multivariate statistical methods and Neural Nets was appropriate. A variety of normalization techniques and classifying methods were tested on this biologically motivated feature set.