Memory maintenance of gradient speech representations is mediated by their expected utility

AbstractLanguage understanding requires listeners compress large amounts of perceptual information into abstract categories. Cues to categories are distributed broadly, with some appearing substantially later. Speech perception is facilitated if gradient sub-categorical representations of input are maintained in memory, allowing optimal cue integration. But maintenance of the high-dimensional signal would tax memory. We hypothesize speech perception balances these pressures by maintaining gradient representations that are expected to facilitate categorization. Two perception experiments tested this. Between participants, an exposure phase manipulated utility of maintenance: in High-Informativity group, later contextual cues were informative; in Low-Informativity group, they never were. A subsequent test phase measured whether participants maintained gradient representations. The Low-Informativity group showed less maintenance than the High-Informativity group (Experiment 1). We then increased task demands, making targets of manipulation non-obvious to participants (Experiment 2) and found similar patterns. These results suggest listeners are capable of allocating memory to gradient representations based on expected utility.

Return to previous page