MUSIC STRUCTURAL SEGMENTATION BY COMBINING HARMONIC AND TIMBRAL INFORMATION

摘要: We propose a novel model for music structural segmentation aiming at combining harmonic and timbral information. use two-level clustering with splitting initialization random turbulence to produce segment labels using chroma MFCC separately as feature. construct score matrix combine from both aspects. Finally Nonnegative Matrix Factorization Maximum Likelihood are applied extract the final labels. By comparing sparseness, our method is capable of automatically determining number types in given song. The pairwise F-measure algorithm can reach 0.63 without rules knowledge, running on 180 Beatles songs. show be easily associated more sophisticated algorithms extended probabilistic models.