Spectral weights for localization and speech-in-speech recognition with spatial separation of talkers on the horizontal plane.

Some previous research has suggested that sound source localization may not rely on the same cues that support the segregation of speech produced by talkers separated in space. The present experiments evaluated spectral weights for the spatial cues underlying these two tasks by filtering stimuli into 1-octave-wide bands and dispersing them on the horizontal plane. Target stimuli were 100-ms bursts of speech-shaped noise or words produced by 24 male and female talkers, and maskers (when present) were sequences of words. For localization in quiet, weights differed depending on the midpoint and band dispersion range, but they were similar for speech and noise stimuli. For bands dispersed between -15° and +15°, weights peaked at 500 and 1000 Hz. Introducing a speech masker changed the magnitude of weights for localization, but not the relative weight by frequency. For speech-in-speech recognition, sequences of masker words produced predominantly informational masking, such that participants had to rely on spatial cues to segregate the target. As for localization, recognition appeared to rely predominantly on spatial cues in the 500- and 1000-Hz bands. Trial-by-trial data suggest that correct word recognition relied on differences in perceived location of target and masker speech for some but not for all participants.