Articulatory Evidence for Interactivity in Speech Production

Abstract

Traditionally, psychologists and linguists have assumed that phonological speech errors result from the substitution of well-formed segments. However, there is growing evidence from acoustic and articulatory analyses of these errors which suggests ac- tivation from competing phonological representations can cascade to articulation. This thesis assumes a cascading model, and investigates further constraints for psy- cholinguistic models of speech production. Two major questions are addressed: whether such a cascading model should include feedback; and whether phonologi- cal representations are still required if articulation is not well-formed. In order to investigate these questions a new method is introduced for the analysis of artic- ulatory data, and its application for analysing EPG and ultrasound recordings is demonstrated. A speech error elicitation experiment is presented in which acoustic and elec- tropalatography (EPG) signals were recorded. A transcription analysis of both data sets tentatively supports a feedback account for the lexical bias effect. Cru- cially, however, the EPG data in conjunction with a perceptual experiment highlight that categorising speech errors is problematic for a cascaded view of production. Therefore, the new analysis technique is used for a reanalysis of the EPG data. This allows us to abandon a view in which each utterance is an error or not. We demon- strate that articulation is more similar to a competing phonological representation when the competitor yields a real word. This pattern firmly establishes evidence for feedback in speech production. Two additional experiments investigate whether phonological representations, in addition to lower-level representations (e.g., features), are required to account for ill-formed speech. In two tongue-twister experiments we demonstrate with both EPG and ultrasound, that articulation is most variable when there is one compet- ing feature, but not when there are two competing features. This pattern is best accounted for in a feedback framework in which feature representations feedback to reinforce phonological representations.Analysing articulation using a technique which does not require the categorisation of responses allows us to investigate the consequences of cascading. It demonstrates that a cascading model of speech production requires feedback between levels of representation and that phonemes should still be represented even if articulation is malformed.