Quantifying the effects of the Indian English phonological system on accent perception using deep neural networks

The perception of accents in speech is influenced by the speaker’s native phonology, but quantifying this influence remains a challenge. This study aims to quantify how Indian English phonology affects the perception of accents in the English spoken by Hindi speakers. A Gated Recurrent Unit (GRU) based neural network model, Phonet [Vásquez-Correa et al., Proc. Interspeech 2019, 549–553 (2019)], is trained on corpora of spoken Indian English (IE) and General American English (AE) to learn the phonological class probabilities of speech segments of both Englishes in a joint vector space. Class probability vector representations are then generated for IE speech from a test corpus annotated with accent ratings by native speakers of AE. We analyze two key contrasts: the labiodental approximant [ʋ], an allophone of the labiovelar approximant [w] in IE, and the retroflex stop [ʈ], compared with its AE counterpart [t]. Euclidean distances between test segments and mean AE/IE baselines are calculated in the joint vector space. A multinomial logistic regression of the distances on the accent ratings shows that segments more distant from the AE baseline correlate with higher odds of strong accents, with segments more distant from the IE baseline showing lower odds. The methods used in this study have potential applications in sociophonetic research and speech acquisition/learning, providing new tools for understanding accented speech.