Discrimination experiments were performed for a change in the spectral shape of noise bands. The subject's task was to discriminate noise bands with a positive spectral slope from those with a negative spectral slope. Thresholds were measured at several bandwidths and center frequencies, as well as for several noise samples. Experiments were performed while roving the overall intensity. At a fixed center frequency of 1 kHz, sensitivity was best for bandwidths of 3–6 semitones (ST). At larger bandwidths, thresholds increased only slowly. At a fixed bandwidth of 1 ST, threshold hardly changed as a function of the center frequency. At a fixed bandwidth of 58 Hz, threshold was lowest near 500–1000 Hz. Model calculations show that the EWAIF model [Feth, Percept. Psychophys. 15, 375–378 (1974)] can account for the present results if the signal's bandwidth does not exceed 1 ST. The IWAIF model [Anantharaman et al., J. Acoust. Soc. Am. 94, 723–729 1993] can account for the present results only if the signal's bandwidth is smaller than 1 ST but larger than about 25 Hz. Results obtained with broadband signals could be described only qualitatively with the multichannel model [Durlach et al., J. Acoust. Soc. Am. 80, 63–72 (1986)]. Then, the model needs the assumption that either the output of the different frequency bands cannot be optimally combined, or that only two bands are used in the discrimination process. The present results are compared with those obtained with two-tone complexes measured under identical conditions [Versfeld and Houtsma, J. Acoust. Soc. Am. 98, 807–816 (1995)].