This talk has two aims: (i) to clarify the role of different cognitive units such as phonemes, contextual variants and words in sound change and (ii) to show how generalised additive mixed models (GAMMs) can be used to analyse data with a rich dynamic and hierarchical structure. The focus is on two phenomena and their interaction: GOOSE-fronting and yod-dropping. Using data from three generations of Derby English speakers, I analyse dynamic formant data and auditory judgments using GAMMs and mixed effects logistic regression. The data show a number of interesting word-specific and category-level patterns, including frequency effects and context-based variation. The results support models of change that incorporate both word- and sound-level cognitive units.

The data set presents a challenge to regression models due to its size and complex hierarchical structure as well as the non-linear nature of formant trajectories. There are multiple measurements representing each trajectory, and multiple trajectories representing each speaker and word, which calls for mixed effects modelling. Moreover, formant trajectories are typically not straight, which means that the assumptions of linear regression models are not justified. While GAMMs offer an attractive solution to these challenges, they are computationally very demanding and they also put a significant burden on the analyst by requiring them to make non-trivial decisions about the structure of their model. The talk provides a detailed account of how these issues were overcome and explains the rationale for the analytic choices made for this project.

Seminar objectives:

- clarify roles of different cognitive units in sound change

- show how generalised additive mixed models can be used to analyse data with a rich dynamic and hierarchical structure

This seminar is presented by Márton Sóskuthy, lecturer in phonetics and phonology in the Department of Language and Linguistic Science at the University of York. http://www-users.york.ac.uk/~ms1341/