Vowel recognition with a time-delay neural network

Abstract

A time-delay neural network (TDNN) architecture is used for speaker independent recognition of the long vowel sounds. A brief introduction to the TDNN architecture and a description of the data used for the simulation are given. Previously published work is extended by training the network with the linear predictive coding (LPC) coefficients of speech along with fast Fourier transform bin energies and by allowing longer and variable length utterances. The training has been performed with multiple speakers using English rather than Japanese speech. With these modifications, 100% recognition of all vowels spoken by two speakers was obtained.