Use of note partials’ energy to improve the identification of polyphonic piano signals

Abstract

The identification of the fundamental frequency of musical sounds is a difficult task which has been studied for more than 20 years. Methods using spectral approaches are used to identify polyphonic sounds. A novel approach for identifying monophonic and polyphonic piano sound signals has been developed by the authors [L. Rossi and G. Girolami, J. Acoust. Soc. Am. 100, 2842(A) (1996)]. The identification is carried out by comparing the incoming spectrum with a database (automatically built previously) containing the frequential distribution of the partials of each piano note. The rate of correct identification is in the 90% range for polyphonic sounds. One kind of error is due to the artificial construction of a note by the combination of partials belonging to other notes; another one results from the difficulty of separating the case of one note played a long time with the case of this same note played two successive times with an overlap. The aim of this paper is to show how the evolution of the partials’ energy associated with the search for information about the time of the note attacks can be used to reduce the false identification.