Abstract

Every native speaker can hear the difference between (English) súbject and subjéct or between (Dutch) voorkómen and vóorkomen. Human listeners use lexical stress for segmentation and disambiguation. However, lexical stress is not normally modelled in automatic continuous speech recognisers. In this work it is modelled how lexical stress can be used in a speech recogniser. Though earlier efforts have not modelled stress for consonants, they appear to contain stress information as well. Furthermore, different spectral features are needed for different phonemes.
A baseline speech recogniser for Dutch and one that uses lexical stress infor- mation are trained. The stress-enabled recogniser's word error rate is lower by 2.6 %.