A frequency domain method for speech separation in a reverberant room

The paper presents an experimental implementation of a speech separation system. Two sources (loudspeakers) and two sensors in a reverberant room are considered. First, the room impulse responses from each source to each sensor are determined. Then, two speech signals are played to the loudspeakers and the two mixtures are recorded by the two sensors. Using the Fourier transforms of the mixtures and room impulse responses, the original speech signals are separated.