Isolated Word Analysis Using Biologically-Based Neural Networks

Abstract

A dynamic synapse neural network (DSNN) for speech recognition system input filtering and the genetic algorithm (GA) used to optimize DSNN parameters is presented. DSNNs are trained to respond to a target word (TW) said by one female speaker or by 8 male and 8 female speakers. The response of the single-speaker trained DSNNs to all 16 speakers is similar to the 16-speaker-trained DSNN responses. TW training results in an ordering of the expected responses to the 9 words of the non-TW set. The ordering determined by single-speaker training matches the ordering determined by multi-speaker training; and in many instances, the single-speaker trained DSNN output matches the multi-speaker trained DSNN output. While searching the parameter space to best solve the isolated word recognition task, the GA implicitly searched the input space to find the input subset best describing the separatrix between TWs and non-TWs. Computation is decreased by concentrating optimization on this subset. The GA adapts as knowledge of this subset is learned. The GA begins as a random search, becoming a steady state GA and then a simple elitist GA over the course of optimization.