A nonwhitening post filter to improve the performance of sound source separation

Abstract

This paper presents a system that provides the capability of humanoid robot to separate sound from the mixture of sounds. The system utilizes a Kulback-Leibler minimization scheme algorithm. A nonwhitening post filter, which is a diagonal inverse of the separation filter, is implemented to improve the separation performance. The results show that a real recording of a speech signal mixed with loud music is well separated and the utilization of nonwhitening post filter improves the performance of separation signal in the frequency domain.