What Computations Support Adaptivity in Human Speech Perception?

-----

Pre-registration required for Zoom via email to thoksber@uci.edu.

---

Prof. Jaeger’s research interest: Models of human speech perception and production; Adaptive changes in language processing; Learning over non-stationary inputs; Theories of communication; Bayesian inference; Advanced multivariate and distributional data analysis (GLMMs, GAMMs, NLMMs, mixture models, etc.)

---

Human speech perception is strikingly adaptive—a response to the infamous “lack of invariance problem”: the mapping from acoustic properties to sound categories and words differs across talkers. On the one hand, this lack of invariance provides a rich source of information for listeners’ inferences about talkers’ physiology (vocal tract size), language background, and social identity. On the other hand, it raises the question how listeners manage to typically understand each other. Why is the “lack of invariance problem” typically not a problem for listeners? We now know that human speech perception seems to solve this problem through adaptation: a single unexpected pronunciation can sometimes be sufficient to change listeners’ perception of subsequent input; even for more complex inputs, like an unfamiliar second language accent, a minute of exposure can yield significant improvements in listeners’ speed and accuracy of understanding that speech. What mechanisms—from low-level auditory transformations to decision-making—support such flexibility? Jaeger will review research in his lab that has tried to address these questions, why and how we try to use model-guided designs and analyses, the challenges we have encountered, some progress we might have made, and what remains to be done.
Specifically, he’ll present evidence that 1) listeners integrate prior expectations with incoming input to facilitate fast but robust adaptation guided by expectations about cross-talker variability, 2) this proceeds through re-weighting of previously learned expectation, rather than unbounded rational information integration, and 3) computationally simpler processes alone—e.g., based on category-independent perceptual compensation or changing decision thresholds/biases—cannot explain the adaptive capacity of human speech perception.

-----