Dynamic chroma feature vectors with applications to cover song identification

Abstract

A new chroma-based dynamic feature vector is proposed inspired by psychophysical observations that the human auditory system detects reltative pitch changes rather than absolute pitch values. The proposed chroma-based dynamic feature vector describes the relative pitch change intervals. The utility of the proposed feature vector incorporated with a music fingerprint extraction algorithm is experimentally explored within a music cover song identification framework. The results with a classical music database suggest that the proposed biologically plausible dynamic chroma feature vector can be successfully added to the conventional chroma feature vector as a complementary feature; it provides a 5.8% relative performance improvement.