Meredith Tamminga, Aaron Ecay, and I have updated code and renamed a paper, Generalized Additive Mixed Models for intraspeaker variation, which has been accepted for publication at Linguistic Vanguard. We suggest GAMMs as a useful tool for separating out different sources of repetitiveness in naturalistic speech data, taking DH-stoping (‘that’ vs. ‘dat’) in the Philadelphia Neighborhood corpus as an example.

For each interview we want to distinguish two potential sources of repetitiveness in the use of the different DH variants like ‘that’ and ‘dat’.

• The tendency of speakers to repeat themselves. Using ‘that’ primes subsequent use of ‘that’.
• The tendency of speakers to use ‘that’ in more formal styles, and ‘dat’ in less formal styles

We expect priming to be a fact about how brains and cognition work, but we don’t have any expectations about how styles change over a given interview. GAMMs allows us to model priming as the effect of the previous token on the current token and model style as a smooth function of elapsed time.

Below is a plot of the random smooths by speaker. Each curve, roughly speaking, represents an estimate of the style for a given speaker in an interview. See the data, code, and paper for more details, but intuitively, the wigglier or more non-linear the smooth, the more varied the style over the course of the interview.