A signal processing method for the separation of concurrent harmonic
sounds is described. The method is based on a two-stage approach.
First, a multipitch estimator is applied to find initial sound
parameters which are reliable, but inaccurate and static. In a
second stage, more accurate and time-varying sinusoidal parameters
are estimated in an iterative procedure, which imposes certain
constraints for the amplitudes and frequencies between the
components of each sound. The proposed algorithm makes it possible
to reconstruct separated sounds that are perceptually close to the
original ones before mixing. Experimental data comprised sung vowels
and 26 musical instruments. Sound separation was performed for
random sound mixtures ranging from one to six simultaneous sounds.
The system is able to produce meaningful results in all polyphonies,
the quality of separated sounds gradually degrading along with the
polyphony.
