Multipitch estimation using judge-based model

Abstract

Multipitch estimation, also known as multiple fundamental frequency (F0) estimation, is an important part of the Music Information Retrieval (MIR) field. Although there have been many different approaches proposed, none of them has ever exceeded the abilities of a trained musician. In this work, an iterative cancellation method is analysed, being applied to three different sound representations - salience spectrum obtained using Constant-Q Transform, cepstrum and enhanced autocorrelation result. Real-life recordings of different musical instruments are used as a database and the parameters of the solution are optimized using a simple yet effective metaheuristic approach - the Luus-Jaakola algorithm. The presented approach results in 85% efficiency on the test database.