Blind speech extraction based on modulus diversity constraint

Abstract

In the speech blind source separation (BSS), the convolutive mixed signals can be obtained when the speech goes through a microphone array. The convolutive mixtures can be solved efficiently in the frequency domain when independent component analysis (ICA) is performed separately in each frequency bin. However, it will cause the problem of permutation ambiguity in each frequency bin. Different from BSS, in the blind signal extraction (BSE) of the speech, we only need to extract the target speech from the noisy environment. Constrained ICA (CICA) is one of the solution solving the BSE problems. Compared with ICA, CICA takes full advantage of the prior information of source signals and the prior information can be used as the additional constraint. In order to solve the BSE problem, this paper presents a frequency domain blind signal extraction method to get the target speech in the presence of diffuse background noise. The proposed method is based on the statistical difference of the modulus between speech and noise in frequency domain where the difference is used as a constrained condition. Results of computer experiments demonstrate the efficiency of the proposed method.