Effectiveness of Signal Segmentation for Music Content Representation

摘要

In this paper we compare the effectiveness of rhythm based signal segmentation technique with the traditional fixed length segmentation for music contents representation. We consider vocal regions, instrumental regions and chords which represent the harmony as different classes of music contents to be represented. The effectiveness of segmentation for music content representation is measured based on intra class feature stability, inter class high feature deviation and class modeling accuracy. Experimental results reveal music content representation is improved with rhythm based signal segmentation than with fixed length segmentation. With rhythm based segmentation, vocal and instrumental modeling accuracy and chord modeling accuracy are improved by 12% and 8% respectively.