In this paper, we propose a noise-robust blind speech separation (BSS) method by using two microphones. We first use modulation domain real and imaginary spectral subtraction (MRISS) to enhance both magnitude and phase spectra of the speech mixture inputs. We then estimate the direction of arrivals (DOAs) of the speech sources and perform time-acoustic-modulation frequency masking to recover the source signals. Our experimental results in five types of noise conditions have showed the superior performance of the proposed method in comparison with the conventional acoustic-domain DOA based separation method.

Index Terms: time-frequency masking, direction of arrival, modulation frequency, blind speech separation