Cognition-inspired Descriptors for Scalable Cover Song Retrieval

Inspired by representations used in music cognition studies and computational musicology, we propose three simple and interpretable descriptors for use in mid- to high-level computational analysis of musical audio and applications in content-based retrieval. We also argue that the task of scalable cover song retrieval is very suitable for the de- velopment of descriptors that effectively capture musical structures at the song level. The performance of the proposed descriptions in a cover song problem is presented. We further demonstrate that, due to the musically-informed nature of the descriptors, an independently established model of stability and variation in covers songs can be integrated to improve performance