Improving overlapping speaker detection using multiple speaker tracking information

Traditionally, multiple speaker tracking consists of two stages, namely, 1) detection of location measurements, followed by 2) a multiple object tracking approach. In general, these two steps are performed separately, and the tracking performance is highly dependent on the measurement detection rate. The performance of the widely used Steered Response Power (SRP)-based measurement detectors, however, drastically decreases in the overlapping speech scenario, where the dominant speaker frequently masks the low-energy speakers. To overcome this problem, we propose an approach that enhances the probabilistic SRP-based measurement detector, using the multiple speaker information obtained in the tracking step. In doing so, this approach tightly couples the two stages, and increases the detection rate of low-energy speakers during overlapping speech segments. Experiments conducted on the AV16.3 corpus showed a significant improvement of the detection and tracking performance, when the proposed approach is integrated into a Kalman-based multiple speaker tracking framework.