On Phase Importance in Parameter Estimation for Single-Channel Source Separation

A single-channel source separation (SCSS) algorithm is targeted to estimate the underlying unknown signals from their single-channel recorded mixture. Current SCSS methods often neglect the phase information in their parameter estimation and use the noisy phase in the signal reconstruction stage. In this paper, we investigate the impact of phase information in the parameter estimation stage of SCSS algorithms and propose using the complete form of the minimum mean square error (MMSE) estimator for mixture magnitude spectrum. We show that previous phase-independent state-of-the-art mixture estimators are special cases of the complete MMSE estimator that takes the phase information into account. Through our experiments, conducted on both synthetic and real signals, we show that the proposed phase-based estimator provides the highest mixture estimation accuracy compared to other phase-independent mixture estimators.