Improved music similarity computation based on tone objects

Abstract

In this paper, we propose a novel approach for music similarity estimation. It combines temporal segmentation of music signals with source separation into so-called tone objects. We solely use the timbre-related audio features Mel- Frequency Cepstral Coefficients (MFCC) and Octave-based Spectral Contrast (OSC) to describe the extracted tone objects. First, we compare our approach to a baseline system that employs frame-wise feature extraction and bagof- frames classification. Second, we set up a system that extracts features on perfectly isolated single track recordings, achieving near perfect classification. Finally, we compare our novel approach against the basis experiments. We find that it clearly outperforms the baseline system in a fiveclass genre classification task. Our results indicate that tone object based feature extraction clearly improves music similarity estimation

The GTZAN dataset: Its contents, its faults, their effects on evaluation, and its future use

“…Note, this compares all tags of an excerpt to only the top tags of label r. Now, if C(X , T g ) is too small (the tags of the excerpt have too little in common with the top tags of its label), or if C(X , T r ) is too large for an r = g (the tags of the excerpt have more in common with the top tags of another label), then we say the excerpt is mislabeled. Creed "I'm Eighteen" (21); Living Colour "Glamour Boys" (29); The New Bomb Turks "Hammerless Nail" (46), "Jukebox Lean" (50), "Jeers of a Clown" (51); Queen "Tie Your Mother Down" (58), "Tear it up" (59), "We Will Rock You" (60); Def Leppard "Pour Some Sugar On Me" (71) Morphine (0-9); Ani DiFranco (10)(11)(12)(13)(14)(15); Queen (16)(17)(18)(19)(20)(21)(22)(23)(24)(25)(26); The Rolling Stones (28)(29)(30)(31)33,35,37); Led Zeppelin (32,(39)(40)(41)(42)(43)(44)(45)(46)(47)(48); Simple Minds (49)(50)(51)…”