Over the past two decades, variation has been promoted from the margins of phonological theory to its center. The success of a phonological model is now measured, inter alia, by how well it accounts for phonological variation. Much progress has been made, and there are currently multiple competing models of phonology, any of which can account for variation with seemingly great success (Anttila 1997, Boersma & Hayes 2001, Coetzee 2006, Kimper 2011, etc.).
In this presentation, however, I will argue that the success of these models is only apparent. All of these models are exclusively grammatical – they do not allow for factors other than grammar to influence the realization of phonological variation. Although phonological grammar undoubtedly contributes to variation, there are many other factors that also influence how variation is realized. In a model where grammar alone accounts perfectly for variation, grammar is therefore doing more than its fair share of the work. Rather than an example of a successful model, this would be an example of a model with a too powerful a grammar.
I will propose an augmented version of the noisy Harmonic Grammar model of phonological variation (Coetzee & Pater 2011). In this augmented version, both grammatical and non-grammatical factors are incorporated, and co-determine how variation is realized. I will rely on two non-grammatical influences on variation as examples: usage frequency and speech rate.
Usage freque ncy. Many variable phonological processes apply more often to frequent than infrequent words. The variable deletion of word-final t/d from consonant clusters in English, for instance, applies more often to frequent just than infrequent jest. The traditional grammatical models of phonological variation do not differentiate between words based on their frequency, and would predict the same deletion rate for both just and jest. The graph to the left serves as an illustration. It plots the t/d-deletion rate in phrase-final position in the Buckeye Corpus (Pitt et al. 2007) against the log frequency of words as measured in CELEX (Baayen et al. 1995). The dotted line marks the prediction of a classic, non-augmented Harmonic Grammar model, and the solid line the prediction of the proposed frequency-augmented model. The augmented model accounts better for the actual deletion rates observed for individual words.
Speech rate. Some variable processes, such as English schwa deletion, apply more frequently at faster than slower speech rates. The word potato is more likely to be realized as p_tato at faster than slower speech rates (Patterson et al. 2003). I will show that listeners use this correlation between speech rate and deletion during speech perception. They are more likely to “perceive” an absent schwa (i.e. to identify p_tato as an utterance of potato) when listening to faster than slower speech. I then develop a model of the listener’s perceptual grammar in the same augmented version of noisy Harmonic Grammar used above, showing how this augmented model accounts better for listeners’ actual performance than a classic non-augmented model.
Given the data sources (speech corpora, data collected in the phonology laboratory, etc.) that we currently have at our disposal, we have access to more realistic information on actual linguistic behavior. The mathematically and computationally sophisticated theories of grammar currently available also enable us to develop more realistic models of grammar. We are at an exciting nexus in the development of linguistic theorizing where we can start integrating traditional generative/competence models with actual production data.