Conceptual blending when used as a creative tool combines the features of two input spaces, generating new blended spaces that share the common structure of the inputs, as well as different combinations of their non-common parts. In the case of music, conceptual blending has been employed creatively, among others, in generating new cadences (pairs of chords that conclude musical phrases). Given a specific set of input cadences together with their blends, this paper addresses the following question: are some musical features of cadences more salient than others in defining perceived relations between input and blended cadences? To this end, behavioural data from a pairwise dissimilarity listening test using input and blended cadences as stimuli were collected, thus allowing the construction of a ‘ground-truth’ human-based perceptual space of cadences. Afterwards, the salience of each cadence feature was adjusted through the Differential Evolution (DE) algorithm, providing a system-perceived space of cadences that optimally matched the ground-truth space. Results in a specific example of cadence blending indicated that some features were distinguishably more salient than others. This pilot study was a first step towards building self-aware blending systems and revealed that the salience of features in conceptual blending is an essential part for producing perceptually relevant blends.