Automatic Differentiation Between Accents of Native and Non-Native English, and the Significance of Prosody

Ladan Baghai-Ravary, University of Oxford

This paper analyses 25 different accents of English, determining differences using both prosodic and non-prosodic features. A Hidden Markov Model aligner phonetically labelled the data. Prosodic features of the phonemes, in the same linguistic and phonetic contexts, were calculated. For non-prosodic comparisons, Dynamic Time Warping (DTW) was used to measure segmental acoustic differences between the phonemes of a given accent with those of other speakers. The discriminative ability of the prosodic features was compared with that of the segmental acoustic score, quantifying the relative utility of prosody and segmental acoustic information in identifying accent.