An Improved Blind Speech Separation Algorithm via Time-frequency Masking Based on a Single Acoustic Vector Sensor

An improved blind speech separation algorithm is presented based on the direction of arrival(DOA) estimation, which is obtained by the precise direction finding ability of a single acoustic vector sensor(AVS). The proposed algorithm works in time-frequency domain, in which the probability at each time-frequency unit of a specific source is estimated via an expectation-maximization(EM) algorithm based on the von Mises distribution mixture model. Because the mean value is difficult to estimate when the reverberation level is high or the sources are placed closely, a simple but effective improved algorithm is proposed, and is verified via simulation under different reverberation level, direction difference and source number. Simulation results show that the improved algorithm is superior to the binary time-frequency masking algorithm and the soft time-frequency masking algorithm in terms of signal-to-distortion ratio(SDR) and perceptual evaluation of speech quality(PESQ)