Noisy vibrotactile signals transmitted during tactile explorations of an object provide precious information on the nature of its surface. Linking the properties of such vibrotactile signals to the way they are interpreted by the haptic sensory system remains challenging. In this study, we investigated humans' perception of noisy, stationary vibrations recorded during exploration of textures and reproduced using a vibrotactile actuator. Since intensity is a well-established essential perceptual attribute, an intensity equalization was first conducted, providing a model for its estimation. The equalized stimuli were further used to identify the most salient spectral features in a second experiment using dissimilarity estimations between pairs of vibrations. Based on dimensionally reduced spectral representations, linear models of dissimilarity prediction showed that the balance between low and high frequencies was the most important cue. Formal validation of this result was achieved through a Mushra experiment, where participants assessed the fidelity of resynthesized vibrations with various distorted frequency balances. These findings offer valuable insights into human vibrotactile perception and establish a computational framework for analyzing vibrations as humans do. Moreover, they pave the way for signal synthesis and compression based on sparse representations, holding significance for applications involving complex vibratory feedback.