Multi-speaker separation employing microphone array and vertex finding algorithm

Abstract

© 2018 IEEE. This paper proposes a new speaker detection and signal separation algorithm for multiple speakers using microphone array data recorded in a room environment. The algorithm utilizes the fact that in multi-speaker conversations not all speakers are speaking simultaneously there are time segments when only a single speaker is active. Based on that observation a speech activity detector for each speaker (MVAD) has been developed. It is based on SRP-PHAT estimates for different blocks of data. We have shown that these estimates form vertexes in a convex polygon which can be employed to obtain MVAD detections. Those detections are then used to form Minimum Variance Distortionless Response (MVDR) beamformers. Evaluations based on real recorded speech data with 4 speakers show that the algorithm provides good interference suppression and low speech distortion for this difficult scenario.