Automatic Extraction of Musical Structure Using Pitch Class Distribution Features

This paper compares the efficiency of different sets of tonal descriptors in music structural discovery. Herein, we analyze the use of three different pitch-class distribution features, i.e. Constant-Q Profile, Pitch Class Profile (PCP) and Harmonic Pitch Class Profile (HPCP), to perform structural analysis of a piece of music audio. We hypothesize that proper segmentation serves as an important basis to obtain music structure analyses of better quality. Thus, we compare the segmentation results produced by each feature to examine its efficiency. A database of 56 audio files (songs by The Beatles) is used for evaluation. In addition, we also show the validity of the descriptors in our structural description system by comparing its segmentation results with a present approach by Chai [1] using the same database. The experimental results show that the HPCP performs best yielding an average of 82% of accuracy in identifying structural boundaries in music audio signals.