Identification of stimulus cues in narrow-band tone-in-noise detection using sparse observer models.

The classical psychophysical paradigm of narrow-band tone-in-noise (TiN) detection has been under investigation for more than 70 years, yet no conclusive answer has been given as to which auditory stimulus features listeners rely on. Here, individual observer models were fit to a large trial-by-trial behavioral data set using a modern statistical analysis procedure. Relative perceptual weights were estimated for a set of auditory features including sound energy, representations of the spectra as well as summary statistics of both fine structure and envelope. The fitted models captured the behavior of all listeners on a single-trial level. The estimated perceptual weights were stable across signal levels. They suggest that responses of observers depended on stimulus energy, though that cue was not always dominant, as well as on band-pass detectors applied to the fine structure spectrum. A subset of the observers exhibited an additional dependence on sound envelope which was best captured by two envelope descriptors: average slope and extrema count. For some listeners, a concurrent analysis of sequential dependencies showed interactions between the current and several preceding decisions. There was no unique answer regarding the strategy individual listeners employ during TiN detection, and implications thereof are discussed.