Entropy per chroma for Cover song identification

Abstract

A Cover song is a rendition of a previously recorded piece of music, cover song identification is about automatically recognizing a song as perceptually the same as another performance of the same song even thought it was played in a different place, perhaps by other musicians each with their own musical instruments. It is a challenging problem of great interest in audio-signal processing, renditions differ in rhythm, tempo, and instrumentation, so it is much more difficult than the classical audio-fingerprinting problem. We extract features from the audio-signal that measure the information content level of the signal grouped by semitones through all octaves, we determine the entropy per chroma value, then use dynamic programming techniques for aligning the renditions since they have different evolution in time and do not last the same. Our tests use 23 performances of two piano pieces and achieved excellent results.