Single Channel Music Source Separation based on Harmonic Structure Estimation

Dongmei Wang, Qinghua Huang

Abstract: Single channel music separation is a useful but difficult problem in audio signal processing field. In this paper a new method is proposed. The method consists of three stages: estimating the harmonic structure of each source in every frame based on iteration of the mixed spectral peaks, clustering the estimated harmonics into the signals they belong to with pitch and formant information, and synthesizing the music source in time domain. Moreover, the method can solve the octave overlapping problem which is a tough one in the single channel source separation area. The experimental results show that our algorithm can separate the mixed signal and obtains a good subjective audio quality.