In current models of modulation perception, the stimuli are first filtered and nonlinearly transformed (mostly half-wave rectified). In order to model the low-pass characteristic of measured modulation transfer functions, the next stage in the models is a first-order low-pass filter with a typical cutoff frequency of 50 to 60 Hz. From physiological studies in mammals it is known that many neurons in, e.g., the inferior colliculus, show a bandpass characteristic in their sensitivity to amplitude modulation. Results from psychophysical studies of modulation masking also suggest some kind of bandpass analysis of modulation frequencies. Results of two experiments on modulation detection that allow discrimination between models incorporating a low-pass filter and those using a modulation filterbank are presented. In the first experiment, modulation detection thresholds were measured for noise carriers of bandwidths between 3 and 6000 Hz. In the second experiment, modulation detection for a sinusoidal carrier was measured in the presence of interfering modulation components with a bandpass characteristic in the modulation spectrum. The results from these experiments could not be simulated by a model including a modulation low-pass filter, but were successfully simulated by a model using a modulation filterbank.