Acoustic convergence to English sibilants is lexically-specific: Evidence for exemplar models

Convergence occurs when talkers alter pronunciation toward speech they hear. Exemplar models provide one account of convergence, as recently heard exemplars are activated during speech production (Goldinger, 1998). Exemplars of speech sounds are stored with their lexical item, predicting that convergence should be lexically-specific. This means that phonetic values should shift towards the model/interlocutor for each word, even if that entails increasing values for some words and decreasing them for others. To test this, a laboratory shadowing study was conducted using English /s/. 34 English speakers produced 80 /s/-initial words (balanced for lexical frequency and following vowel rounding) with fillers, then were passively exposed to model speech with the same /s/-initial words produced by a female talker, then produced the words again post-exposure. Spectral center of gravity (CoG) for /s/ was extracted using Praat (Boersma, 2001). Most participants produced baseline CoG values that were both lower and higher than the model for different words. When the data are analyzed by baseline relative to the model, a significant convergent pattern emerges: words initially produced with higher CoG shifted down after exposure and words initially produced lower shifted up. Participants converged to the model's acoustic values for CoG and direction of shift was lexically specific. This indicates that talker- and word-specific acoustic variability is stored and utilized, as predicted by exemplar models of speech sound representation.