Collecting and analyzing spoken utterances for a speech controlled application

Abstract

To estimate the parameters of stochastic knowledge bases for a speech understanding system, many utterances spoken by many people have to be examined. For the regarded domain of a 'graphic editor', two different manners of collecting training data are discussed. An analysis of spoken commands recorded by a 'Wizard of Oz'-simulation shows that the way to talk to a computer usually depends on how familiar a subject is with a computer.