Using hidden Markov models to find discrete targets in continuous sociophonetic data

Abstract

Abstract Advances in sociophonetic research resulted in features once sorted into discrete bins now being measured continuously. This has implied a shift in what sociolinguists view as the abstract representation of the sociolinguistic variable. When measured discretely, variation is variation in selection: one variant is selected for production, and factors influencing language variation and change are influencing the frequency at which variants are selected. Measured continuously, variation is variation in execution: speakers have a single target for production, which they approximate with varying success. This paper suggests that both approaches can and should be considered in sociophonetic analysis. To that end, I offer the use of hidden Markov models (HMMs) as a novel approach to find speakers’ multiple targets within continuous data. Using the lot vowel among whites in Greater St. Louis as a case study, I compare 2-state and 1-state HMMs constructed at the individual speaker level. Ten of fifty-two speakers’ production is shown to involve the regular use of distinct fronted and backed variants of the vowel. This finding illustrates HMMs’ capacity to allow us to consider variation as both variant selection and execution, making them a useful tool in the analysis of sociophonetic data.