Anticipated ITD statistics built-in human sound localization

The variability of natural scenes places perceptual processes in the realm of statistical inference, where sensory evidence must be weighted by its reliability. Absent prior information, estimating environmental variability requires real-time sampling and computations. However, a portion of environmental variability can be assumed invariant across conditions. Perceptual tasks relying on time-dependent information may be vastly enhanced if the invariant statistical structure of sensory cues is built into the underlying neural processing. We investigated this question in human sound localization, where the statistics of spatial cues can be estimated. Localizing low frequency sounds in the horizontal plane relies on interaural time differences (ITD). We estimated the ITD statistics across frequency and azimuth from human head-related transfer functions (HRTFs). The mean ITD varied with azimuth following a sigmoid relationship, whose slope is steepest at the center. In addition, ITD was more variable over time for sounds located in the periphery compared to the center, in a frequency-dependent manner. We investigated the role of these statistics -- ITD slope and ITD variability -- in low-frequency lateralization of human subjects, to test the hypothesis that high-order sensory statistics are represented in the human brain influencing spatial discriminability and novelty detection. Thresholds for discriminating ITD changes reported by classical studies (Mills, 1958) were predicted by a model that considered both ITD slope and ITD variability. To further test our hypothesis, EEG novelty responses were recorded in human subjects undergoing an oddball stimulation sequence, where repetitive ("standard") tones of a given ITD were combined with sporadic ("deviant") tones of a different ITD. By using insert earphones, ITD was shifted with zero variability across time and location. Mismatch negativity (MMN) brain signals were used as an index of discriminability between standard and deviant stimuli. We found that MMNs were weaker for standards in the periphery, where the ITD slope is lower and the ITD variability is higher. Overall, the amplitude of novelty EEG signals was predicted by the difference in ITD between the standard and deviant normalized by the anticipated discriminability of the standard location, indicating that change detection is weighted by expected statistics of the sensory input. In sum, our results show that spatial discriminability thresholds and deviant detection are consistent with a representation of anticipated ITD statistics in the human brain, supporting the hypothesis that high-order statistics are built into human perceptual processes biasing behavior.