Speech Science Forum - Bodo Winter (University of Birmingham)

Abstract:

Linguistics and phonetics have undergone many quantitative revolutions that deeply affected the way data analysis is conducted, such as the shift away from significance tests like t-tests and ANOVAs to the generalized linear model framework. However, despite massive progress in almost all subfields of the language sciences, certain issues persist, and some age-old issues resurface again and again when new methods are introduced into the field. In this programmatic and discursive talk I hope to excite debate. Inspired by Tal Yarkoni’s recent “generalizability crisis” target article, I will discuss a set of examples relating to both speech production and speech perception experiments all of which surround the core question: what are the sources of variation that we wish to generalize over? Guided by this question, we will look at how particular methodological paradigms may be ignoring theoretically relevant sources of variation either in experimental design or in statistical methodology. I will also end with a call towards incorporating more Bayesian analysis into phonetics. While Bayesian models are no panacea, they present an opportunity for the field to step outside of deeply engrained statistical rituals.