TOWARDS PHRASE STRUCTURE RECONSTRUCTION FROM EXPRESSIVE PERFORMANCE DATA

Using a simple pattern finding approach, we investigate to what degree patterns found in the tempo and loudness curves measured from piano performances of a classical piece coincide with repeated musical structures in the score that was performed. We show that high frequency content in such curves is more useful for finding repetitions of musical structures than low frequency content. In some cases removing low frequency content even improves the accuracy of pattern finding. In this paper we investigate to what degree the phrase structure of a piece is reflected in the tempo and loudness information measured from performances of the piece. We do this by measuring how well patterns found in the tempo and loudness curves coincide with the phrase structure, more specifically melodic gestures, relatively small musical constructs (typically containing less than ten notes). Rather than determining the precise beginnings and endings of phrases and melodic gestures, our first goal is to establish which parts of the piece are repeated, and where. We measure accuracy in terms of how many of the instances of the pattern span repeated melodic gestures (precision), and how many repeated melodic gestures are identified as instances of the same pattern (recall). Obviously, a phrase structure reconstruction is not correct if the boundaries of the melodic gestures are not correct, but we believe that if repeated parts of the score are identified largely correctly, a useful step towards phrase structure reconstruction has been made.