Against listener-oriented sub-phonemic differentiation

Abstract

The English word "like" has attracted much sociolinguistic interest, but also phonetic and phonological research: Due to its many functions (Podlubny et al. 2015 report 11 in Canadian English, but we argue there may be as many as 16), it is a prime target for studies of sub-phonemic differences between near-homophones. For instance, Drager (2009) famously showed differences in segment length or realisation for different functions of "like" in New Zealand English; Podlubny et al. (2015) similarly found vowel realisation and length differences between "like" functions in Canadian English. However, these differences are small, when they are found at all – Schleef and Turton (2018) do find different vowel realisations between "like" functions in Edinburgh and London varieties of English, and argue that these are due only to prosodic contexts for certain functions favouring reduction. This possibility casts doubt on how systematic and thus how transmittable/learnable such differences would be. Therefore, we study "like" in another regional accent (or pan-regional standard, following Strycharczuk et al. 2020) in England to see if differences exist there; the questions of origin and transmission routes would, of course, be for future research.
We recorded 11 young adult (age range: 18 to 25 years) speakers of English from the North-West of England in informal conversation with a family member or friend (as in e.g. Warner and Tucker 2011), and also reading a list of 36 sentences containing different functions of like. The conversations were transcribed manually; transcripts and sentence-lists were force-aligned to recordings using the self-training Montreal Forced Aligner (McAuliffe et al. 2019). We extracted all "like" tokens and calculated/annotated segmental and word-level features (namely the duration of every token and segment, average speech rate in a window extending up to 3 words either side of the token, F1 and F2 at 25% and 75% through each vowel segment, and the Euclidean distance between these formant values as a measure of diphthongisation) as well as context features (Beckman and Hirschberg 1994's ToBI break index strength following the token, position of the "like" token in the utterance, and the segments and words immediately preceding and following the token). To account for predictability effects on pronunciation (e.g. Hall et al. 2018), we extracted the bigram frequencies either side from SUBTLEX (van Heuven et al. 2014). We used mixed-effects regression models and agglomerative hierarchical clustering to investigate this data for any systematic differences.
Counter to prior research, we find no systematic acoustic differences between "like"s of different functions: Four separate regression models (with the word length, /k/ segment length, ratio of /l/ segment length to vowel segment length, and the diphthongisation measure as dependent variables respectively) as well as hierarchical clustering all fail to show any reliable difference in like realisation by "like" function. The only strong acoustic differences we find are between male and female speakers (in pitch and formants) as well as between conversation and sentence-list tokens (longer tokens and more diphthongal vowels in sentence-list reading).
The sex and genre differences are unsurprising, but serve as sanity checks. The fact that we found no other reliable differences in like realisation by function shows that the North-West England accent does not differentiate between functions of "like" phonetically, despite how useful this would be given the number of functions. This, we argue, suggests that listener-oriented accounts of different mental representations for (near-)homophones are not borne out.