Memory for repeated auditory textures

Even though memory plays a pervasive role in perception, the nature of the memory traces left by past sounds is still largely mysterious. Here, we probed the memory for natural auditory textures. For such stochastic sounds, two types of representations have been put forward: a representation based on sets of temporally local features, or a representation based on time-averaged summary statistics. We synthesized naturalistic texture exemplars and used them in an implicit memory paradigm based on repetition, previously shown to induce rapid learning for artificial sounds such as white noise. Results were similar for artificial and natural sounds, exhibiting a general trend for a decrease in repetition detection performance with increasing exemplar duration, although with some variation depending on texture type. This trend could be captured by a summary statistics model, but also by a new model based on the random sampling of temporally local features. Moreover, repeated exposure to a same natural texture or artificial noise exemplar systematically induced a performance gain, which was comparable across all sound types and exemplar durations. Thus, natural texture exemplars were amenable to learning when repeated exposure was available. The findings are consistent with two interpretations: the existence of a special processing mode when acoustic repetition is involved, to which natural textures are not immune, or a convergence of the local features versus summary statistics descriptions if a continuum of time scales is considered for auditory representations.