Analysis and improvement of post-processing for BSS aiming to suppress recognizable crosstalk

The performance of the blind signal separation (BSS) must be evaluated using not only quality of separated signals but also absolute level of residual noise signals (crosstalk components). The purpose of our study is to suppress the recognizable crosstalk components and to extract only the target speech using a post-processing unit. In this paper, the proposed methods are analyzed in detail. Additionally, the efficient post-processing is newly proposed. The experimental results show that the proposed method achieves preventing the recognition of the crosstalk components without sacrificing hearing of the target speech.