Single-channel speaker separation based on sub-spectrum GMM and Bayesian theory

The problem of single-channel speaker separation attempts to extract the speech signal uttered by the speaker of interest from one channel signals containing a mixture of acoustic signals. Most of current techniques failed to eliminate the interfering signal completely. In this paper, we present a new approach to solve this problem. Itpsilas an iterative separation approach based on sub-spectrum GMM and Bayesian theory. First, we obtain sub-spectrum GMM models for each speaker in the training phase. Then, separated speech signals are estimated based on Bayesian model given the mixture. Finally, an iterative separation process is used to separate out the speech signal of the speaker of interest from the mixture. Simulation results exhibit a high level of separating performance.