Stable and Low-Distortion Algorithm Based on Overdetermined Blind Separation for Convolutive Mixtures of Speech

Abstract

We propose a new algorithm with a stable learning and low-distortion based on overdetermined blind separation for the convolutive mixture of the speech. To improve the separation performance, we have proposed multistage ICA, in which frequency-domain ICA and time-domain ICA (TDICA) are cascaded. For temporally correlated signals, we must use TDICA with a nonholonomic constraint to avoid the decorrelation effect. However, the stability cannot be guaranteed in the nonholonomic case. Also, in the holonomic case, the sound quality of the separated signal is distorted by the decorrelation effect. To solve the problem of the stability, we perform TDICA with the holonomic constraint. To avoid the distortions, we estimate the distortion components by TDICA with the holonomic constraint and we compensate the sound qualities by using the estimated components. The stability of the proposed algorithm can be guaranteed by the holonomic constraint, and the proposed compensation work prevents the distortion. The experiments in a reverberant room reveal that the algorithm results in higher stability and higher separation performance.