Extracting refrained phrases from music signals using a frequent episode pattern mining algorithm

In this paper, we discuss a method for extracting refrained phrases from a music signal by a discrete knowledge discovery processing approach instead of a signal processing approach. The proposed method consists of two processes: translating a music signal into a sequence of events that represent pitch information, and then mining the frequent patterns from the event sequences. The former is performed by computing chroma vectors at every beat interval, and the latter is performed by enumerating the frequent episode patterns. We carried out a preliminary experiment on some pieces in the RWC music databases to examine if the extracted patterns represent the refrained phrases.