<p>Vibrotactile pitch perception is shaped by mechanical receptors. However, the effects of the frequency selectivity of these peripheral neurons on the vibrotactile pitch perception is not examinable in biological organisms or in psychophysical experiments. In this study, we develop a hybrid model composed of the mechanoreceptive afferent (MA) model and a deep neural network (DNN), aiming to explain the effects of receptors filtering to mechanical vibrations on the vibrotactile pitch perception. We first train the network to predict the fundamental frequency (F0) from the natural representations of receptor neurons. The network with MA models reproduces a compound perception of the amplitude and frequency in vibrations, which is considered as a typical human-like behavior. Although the one-dimension convolution (Conv1d) kernels in another DNN instead of MA models also learn to bandpass filter vibrations and exhibit the similar sensitive frequency distribution with MA models, the DNN fails to shift the frequency perception when the amplitudes of vibrations are changed. This result indicates that the frequency selectivity in encoders actually contributes to estimate the F0 from natural vibrations as both two networks realize high-accurate F0 classification, but the pitch perception still requires the mechanoreceptive afferent representations. Further, we compare the accuracies of DNNs with MA models in the F0 prediction task and F0 shifting with the amplitudes task, and prove that a better F0 estimation really contributes to the compound pitch perception of vibration frequency and amplitude.</p>