The contribution of salient localizable glimpses on speech intelligibility in a multitalker setting with spatially diffuse signals

Speech intelligibility in a multitalker scene is a challenging task, even for normal-hearing listeners. Spatial separation of the various talkers improves speech intelligibility, leading to a spatial benefit relative to the situation in which all speech signals originate from the same direction.

Typically, in a setting with a small number of talkers, the sparse spectro-temporal character of speech causes interfering speech to only partly mask the target speech. As a result, the local signal-to-noise ratio (SNR) fluctuates strongly within a spectro-temporal representation of simultaneous speech signals.

In a previous experiment (Adv. Exp. Med. Biol. 894, p. 73–81.) spectro-temporal elements of target speech were eliminated from a three-talker mixture when they were lower than a predefined SNR criterion. Measurements of speech intelligibility over a range of such SNR criteria demonstrated that when target speech was reduced to only few, but very salient, spectro-temporal elements („glimpses“) the spatial benefit was sustained and of comparable size to that of intact speech. This suggests that the spatial information contained in high-SNR glimpses is highly important for a proper organization of the auditory scene by the listener.

In a reverberant environment, however, speech signals become interaurally decorrelated, rendering spatial cues less useful as a cue for source segregation. In a new experiment we investigate whether the adverse effect of decorrelated interaural cues can be reduced by delivering useful spatial information to only the most salient glimpses of either the target or the interfering talkers, thereby rendering some of the glimpses localizable.

Our experimental data will be discussed in relation to contributions of monaural and binaural cues to auditory grouping and segregation. In addition, the data will provide insight into the importance of localizability of target versus interfering signals.