An analysis of “speech glimpses” in realistic environments

A great deal of effort is currently going into recreating everyday acoustic environments in laboratories and clinics, with the goal of obtaining more relevant measurements of real-world listening abilities and intervention benefits. Related work is focused on generating naturalistic speech stimuli that capture important features of real conversations, and on estimating real-world signal-to-noise ratios. Here we make use of a framework that brings together all of these approaches to arrive at highly realistic speech-in-noise stimuli. Using ideal time-frequency segregation, we characterized the “speech glimpses” that are available in these real-world stimuli. By speech glimpses, we refer to distributed time-frequency regions in which the speech of interest dominates the acoustic mixture. One goal was to compare the glimpses that are available in highly realistic stimuli to those available in simpler, commonly used, laboratory stimuli. A second goal was to analyze these glimpses in detail, to provide a new perspective on the many sources of disruption that may hinder the understanding of conversational speech. These include masking (which reduces the number of available glimpses) and reverberation (which reduces the quality of glimpses), as well as interactions between sounds that cause distortions of the spatial information present in the target glimpses.