PDOA based underdetermined blind source separation using two microphones

This paper builds on the idea of frequency bin-wise separation of mixed speech signals by Gaussian Mixture Model fitting using the Expectation-Maximization algorithm on Phase-Difference-of-Arrival values between two microphones. We find that using a combination of pre-processing steps and error finding post processing, separation performance exceeding the state of the art is achieved on publicly available live recording data.