Pronunciation variation modelling using accent features

In this paper, we propose a novel method for modelling native accented speech. As an alternative to the notion of dialect, we work with the lower level phonological components of accents, which we term accent features. This provides us with a better understanding of how pronunciation varies and it allows us to give a much more detailed picture of a person’s speech. The accent features are included during phonological adaptation of a speaker-independent Automatic Speech Recognition system in an attempt to make it more robust when exposed to pronunciation variation thus improving recognition performance on accented speech. We employ a dynamic set-up in which the system first identifies the phonetic characteristics of the user’s speech. It then creates a model of the speaker’s phonological system and adapts the pronunciation dictionary to best match his/her speech. Recognition is subsequently carried out using the adapted pronunciation dictionary. Experiments on British English speech data show a significant relative improvement in error rate of 20% compared with the traditional non-adaptive method.