Twelve reproducible noises were used as stimuli in a 2-interval forced-choice signal-detection experiment. The noises were stored numerically in a PB250 computer that converted thein to sound during the experiment by means of a digital-to-analog converter. The 240 numbers specifying a noise were sampled at a rate of 2500 numbers/sec, generating a 96-msec stimulus. A sinusoidal signal with a period of 8 samples/cycle was added to one noise on half the trials. Spectral analyses were computed for all stimuli. On nonsignal trials, biases to particular noises were found that could be explained in part, but not entirely, by differences between the noise pairs in energy around the signal frequency. Performance on signal trials was related to the energy difference between the stimuli in the region near the signal frequency, but was not entirely accounted for by this variable. Special characteristics of certain noises appear to affect the subject's response when these noises appear in either signal or nonsignal trials. Except for bias effects, detection on signal trials with the same noise in both intervals was similar to that on trials with different noises. Implications for theories of signal detection are discussed.