Measuring the Impact of Segmental Deviation on Perceptions of Accentedness using Gradient Phonological Class Features

Using Phonet (Vásquez-Correa et al., 2019), a neural network-based model, we generate vector representations of speech segments consisting of phonological class probabilities and use these representations to quantify segmental deviations in the English of native Hindi speakers from American English (AE) and Indian English (IE) baselines, in order to explain how these deviations impact perceptions of accentedness by native AE speakers. The primary focus is on three AE phonemes and their realizations in Hindi English (HE) and Indian English: the labiovelar approximant /w/, often produced as the labiodental approximant [ʋ]; the alveolar stop /t/, commonly realized as the retroflex stop [ʈ]; and the rhotic approximant /ɹ/,rendered as the rhotic tap [ɾ]. Multinomial logistic regressions of Euclidean distances from HE sements to AE/IE baselines on accent ratings show that larger distances from AE baselines increase the likelihood of perceiving stronger accents while larger distances from IE baselines decrease the likelihood. Changes in the probability distributions of contrastive phonological classes are found to correlate with the strength of the perceived accent. These results offer valuable insights into the interplay between native phonology and the perception of accented speech.