Polyphonic Pitch Extraction from Musical Signals

This paper presents a polyphonic pitch analysis method directly applicable to musical signals. After an initial time-frequency analysis step, the method iteratively projects the spectrum on elementary spectrum models. For each possible perceived pitch, several models are available corresponding to different spectral distributions leading to the same perceived pitch. The results of the method are illustrated on a wide variety of sound signals, and its limits are discussed. The method's appeal lies in its applicability to a wide range of physical or electronic instrument sounds, its being intrinsically designed for polyphonic contexts, its relatively low computational complexity, and its potential for analysis/re-synthesis.