Abstract
The paper presents a new method of forming expert modules for modular time delay neural networks (modular TDNNs) using ensembles of similarly trained TDNNs referred to as squads. Squad base expert modules for closing diphthong recognition are compared with traditional expert modules comprising individual TDNNs and are found to afford significantly better false positive error performances, while recognition performances are comparable or better. Traditional and squad based expert modules formed from three different TDNN variants are compared. One of these variants, sequence token TDNN, embodies a novel method of using traditional TDNNs for closing diphthong recognition and is found to outperform the other variants when squad based expert modules are used.

This publication has 6 references indexed in Scilit: