Blind source separation based on local generalized Gaussian mixture model

In this paper, the local generalized Gaussian mixture model (LGGMM) is proposed to blindly separate speech signals in reverberation environment. First, time delay and attenuation ratio of speech signal to the microphone array are estimated by detection of single source points. Second, by using LGGMM to determine the dominant sources at every time-frequency point, the speech signals are separated. Experiment simulation results show that this method can further improve judgment accuracy of the dominant sources and can better separate the speech signals even if reverberation time is longer.