Date of Completion

8-3-2020

Embargo Period

8-3-2020

Keywords

Speech perception, Psycholinguistics, Distributional learning, Statistical learning

Major Advisor

Rachel M. Theodore

Associate Advisor

Emily B. Myers

Associate Advisor

James S. Magnuson

Associate Advisor

Gerald T. M. Altmann

Field of Study

Speech, Language, and Hearing Sciences

Degree

Doctor of Philosophy

Open Access

Open Access

Abstract

There is no one-to-one mapping between speech acoustics and individual speech sounds. The acoustic cues produced for individual speech sounds show wide variability both within and across talkers. Nonetheless, listeners perceive the speech of familiar and novel talkers with ease. It is theorized that listeners achieve this by maintaining a degree of flexibility in how acoustics are mapped to speech sound categories. Building on previous work demonstrating that listeners are sensitive to individual talker differences in speech production, we test the hypothesis that distributional learning for input statistics is contextually governed by talker identity. Listeners (n = 320) completed two blocks of phonetic identification for VOT input distributions specifying the /g/ and /k/ categories. Input distributions in one block contained relatively shorter VOTs compared to the other. Across listener groups, block order and talker constancy was manipulated. Predictions for talker-specific vs. talker-agnostic distributional learning were derived through simulations performed with the Bayesian belief-updating model of speech adaptation, which yielded qualitatively different patterns of learning for the same-talker vs. different-talker simulations. The results showed (1) robust evidence of distributional learning in that listeners’ voicing boundaries moved block-to-block in line with changes in the input statistics, (2) no difference between the same-talker and different-talker conditions, and (3) learning patterns that were consistent with cumulative integration of distributional input statistics across blocks. Collectively, the results suggest that distributional learning in this paradigm is not talker-specific, which may reflect a-priori informativeness of VOT as a cue to talker identity.

Download
COinS