Blind separation method of overlapped speech mixtures in STFT domain with noise and residual crosstalk suppression

Abstract

Noise and residual crosstalk are two important issues that have to be addressed in practical applications of underdetermined blind source separation (UBSS) for speech mixture. This paper proposes a noise-robust UBSS algorithm to deal with highly overlapped speech sources with residual crosstalk suppression scheme in the short-time Fourier transform (STFT) domain. The proposed algorithm is firstly to estimate the mixture matrix of noisy sources and then the original sources are recovered by suppressing the crosstalk. To reduce the noise effect on the detection of auto-source points in the STFT spectrum, we propose a method to effectively detect the auto-term locations of the sources by using the principal component analysis (PCA) on the STFTs of noisy mixtures. To mitigate the crosstalk of separated speech sources, a Gaussian mixture model (GMM) is implemented. Simulation results show that substantial performance improvement has been achieved.