To separate a single-channel mixture of multiple source signals, it is important to estimate independent features of each source. There are many feature estimation methods which focus on the magnitude spectrogram. On the other hand, features from the phase spectrogram have not yet been fully exploited. In this report, we try to reveal a basic aspect of the phase components by studying iterative phase estimation algorithms in which the magnitude spectrograms of the unknown source signals are assumed to be known. In addition to the consistency of the magnitude spectrograms during the iteration of the short-time Fourier transform and its inversion, our proposed phase estimation algorithm takes into account of the mixture consistency in the spectrogram domain. The performances of the proposed algorithm are demonstrated through some simulation results.
NETL
