An experiment was conducted to measure the shape of the psychometric function for the discrimination of spectral variance. The stimuli were simultaneous tone complexes comprised of the six octave frequencies from 250 to 8000 Hz. On each presentation the levels of components in dB were drawn independently and at random from one of two normal distributions having identical means but different variances (sigma N = 1 dB, sigma S = 2-10 dB). In the standard two-interval, forced-choice procedure, the listeners' task was to indicate which complex had the greater variance in component level. The shape of the psychometric function for all five listeners was markedly different from that of an observer limited only by additive internal noise. It was consistent with an observer that gives weight to only one or two components in the complex. However, this result was inconsistent with the weighting functions computed from the trial-by-trial data from these listeners. Both measures can be reconciled if it is assumed that listener weights vary from trial to trial, or that decisions are based on the one tone in the complex having the maximum level.