The role of learned selective attention in speech adaptation

No two people speak exactly alike. As a result, one person’s "pole" might sound like another person’s "bowl." The difference between such minimally different words comes down to small details in how they are pronounced. Some of these details vary across accents. However, no matter who the speaker is, listeners can quickly adjust to novel accents. This project examines how people adapt to talker differences in speech production and tests the hypothesis that people adapt by shifting their attention to different acoustic cues when it improves accuracy. The research is guided by a computer model that predicts how this process works. Societal benefits include a training app, training for graduate and undergraduate students, and publicly sharing available materials. This project combines computational modeling and behavioral experimentation to test key predictions of a computational model of accent adaptation. The model is built on the ideas of learned selective attention theory and reinforcement learning, representing a novel application of these theories to speech perception. The central hypothesis of the model is that listeners use reinforcement learning from prediction error to rapidly reallocate attention across perceptual dimensions. This project tests (1) whether focusing attention on a perceptual dimension makes it easier to learn about that dimension’s relationship to speech categories, (2) whether individuals who weight a dimension highly are the individuals who find it easiest to change the weight of that dimension, and (3) whether variability along a dimension has a different effect on learning rate depending on whether the dimension is to be up-weighted or down-weighted. In doing so, this project advances an account of learned selective attention theory for speech perception. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.