Description:

David Kleinschmidt
University of Rochester

Coping with variability: Speech perception and beyond

Variability across situations can make perception extremely challenging. This is especially true in speech perception and language comprehension, where the robustness of speech perception in the face of variability across talkers (the "lack of invariance") is a long-standing puzzle. At the same time, variability also provides important information about the current environment, including talker identity and other indexical features like gender, age, dialect, etc. I will present an approach to variability in speech perception and language comprehension that I call the "ideal adapter", which treats speech perception as a problem of inference under uncertainty at different levels. First, listeners need to infer _what_ a talker is trying to say. Second, at the same time, they need to infer _how_ the talker produces language overall, or the distributions of cues that are produced for each underlying linguistic unit. Third, listeners need to infer _who_ the current talker is, relative to other talkers they have encountered, in order to take advantage of prior experience, either with the particular talker, or other similar talkers, or even the range of talkers in the language as a whole. This approach ties together a range of strategies that listeners use to cope with variability across talkers, and opens up new directions for future work.

Finally, I will present ongoing neuroimaging work that probes how the underlying neural representations of perceptual information change in response to changing sensory statistics and task demands. This work highlights parallels between speech perception and perception in other domains. It also demonstrates that _representational similarity analysis_ can be a powerful, generally applicable method for testing specific hypotheses about possible neural representations across different perceptual domains. Future work will apply these methods to understand the neural mechanisms that support robust speech perception.