Strong evidence for maintenance of gradient representations during language processing

To what degree listeners can maintain gradient subcategorical information about speech input in memory over time has been a matter of considerable debate. The literature has largely lacked formal computational models of potential mechanisms against which to compare human behavior. Here, we formalize several competing cognitive models of this process and quantitatively compare them to data from a series of behavioral experiments. We find consistently strong evidence in favor of models which allow for maintenance of subcategorical information over the course of an utterance. These results suggests that listeners are able to maintain relatively fine-grained details about prior linguistic input over long perceptual timescales. This work also highlights the importance of formalizing cognitive models of behavior to distinguish between competing theoretical mechanisms.