Multiple Speech Source Separation by Using MVDR for B-Format Recordings.

This paper presents a multiple speech source separation method that adopts the minimum variance distortionless response (MVDR) beamformer to separate speech sources based on the direction-of-arrival (DOA) estimated from the B-format microphone recordings. Unlike conventional microphone arrays, the B-format microphone provides complete spatial information of the pressure gradient and acoustic pressure which is useful to multiple sources localization. Therefore, this paper adopts the single-source points detection of the B-format microphone recordings, where only the dominant source is present in the time-frequency domain. A higher accuracy of source localization is achieved by identifying and utilizing only single-source points which contain correct direction cues. According to the DOA estimates, MVDR beamformer is then adopted for recovering the sources separately. The objective evaluation shows that the proposed method outperforms the existing BSS algorithm.