Friday, April 25, 2008, 12-2 p.m.

Linking grammatical principles and speech production data

Grammars allow for a precise specification of distinctions between more and less well-formed linguistic structures. For example, a syntactic generative grammar will generate all and only those well-formed sentences of a language (e.g., producing "the dog walks" but not "walks dog the"). Many researchers (e.g., Jakobson, 1941) have proposed that well-formedness plays a critical role not only in the structure of languages but also in their perception, production and acquisition. In this talk, I'll develop such an account in the context of speech production. I'll review empirical evidence from aphasia and experimentally-induced speech errors showing that both categorical and gradient well-formedness distinctions influence speech production. I'll then discuss how these probabilistic patterns can be quantitatively modeled. The Harmonic Grammar formalism is used to precisely specify phonological well-formedness. Specific Harmonic Grammars are then realized within connectionist networks which are subject to stochastic damage processes (resulting in errors). This error generation process accounts for broad patterns of speech error distributions. Further extensions allow quantitative predictions for specific error types to be derived.