A Study on Musical Features for Melody Databases

Abstract

The design of content-based music retrieval systems on the Web is a challenge, since music is auditory, temporal, and multidimensional — the same piece can be interpreted in multiple ways. Most literatures on music retrieval simply map the problem to existing information retrieval paradigms, mainly that of text, by modeling music as a sequence of features. However, this mapping raises questions to be an- swered. Through the study of the statistical properties of six features, namely Profile, Note Duration Ratio Sequence, Interval Sequence and their variants, we answer four of these questions in this paper. They are: the number of musical “alphabets” and “words” in musical features, whether Zipf’s law holds for musical features, whether there are any musical “stopwords”, and the range of n for n-gram based music indices.