Discrimination of frequency variance for tonal sequences

Real-world auditory stimuli are highly variable across occurrences and sources. The present study examined the sensitivity of human listeners to differences in global stimulus variability. In a two-interval, forced-choice task, variance discrimination was measured using sequences of five 100-ms tone pulses. The frequency of each pulse was sampled randomly from a distribution that was Gaussian in logarithmic frequency. In the non-signal interval, the sampled distribution had a variance of σ_STAN², while in the signal interval, the variance of the sequence was σ_SIG² (with σ_SIG² > σ_STAN²). The listener's task was to choose the interval with the larger variance. To constrain possible decision strategies, the mean frequency of the sampling distribution of each interval was randomly chosen for each presentation. Psychometric functions were measured for various values of σ_STAN². Although the performance was remarkably similar across listeners, overall performance was poorer than that of an ideal observer (IO) which perfectly compares interval variances. However, like the IO, Weber's Law behavior was observed, with a constant ratio of (σ_SIG²- σ_STAN²) to σ_STAN² yielding similar performance. A model which degraded the IO with a frequency-resolution noise and a computational noise provided a reasonable fit to the real data.