The Design and Performance of Two Accent Diagnostic 'Shibboleth' Sentences

Speech recognition systems which claim to be speaker independent must either address, or somehow bypass, the problems of inter-speaker variability.The issue may be avoided altogether with the simplest of recognition tasks, but they must be accounted for as complexity increases.Variability results from physical differences in our speech organs, and in the way we use them.We may be categorised by our typical speech style or range, or by a favourite phrase, but the most commonly used identi er of this sort is our regional accent.Although different accents do not cause major comprehension dif culties for speakers of the same language, simple HMM recognition strategies may not be able to cope with all accents and a system for accurately automatically classifying regional accent may be desirable.This paper presents criteria for the diagnosis of nine accents of British English, based on the analysis of Shibboleth speech data.Some initial results of this approach are presented, including suggestions towards an automatic system {or accent decision-making.The speech used to test this strategy came from the Subscriber database, collected by BT Laboratories and annotated at CSTR as part of a collaborative research project.