Adaptive cubature Kalman filtering for distant speech tracking using a circular microphone array

Abstract

A joint processing framework for distant speech tracking is proposed based on combination of an adaptive cubature Kalman Filter with a 80-element circular microphone array. The deconvolved convention beamforming (DCB) [T. C. Yang, 10.1109/JOE.2017.2680818] is carried out over the distant speech data for speech enhancement and bearing estimate. DCB provides superdirective beams and offers the same robustness as conventional beamforming. With the estimated speech bearing, a “current” statistical (CS) motion model is combined with a cubature Kalman filter to track the speech of interest. The CS model assumes a modified Rayleigh distribution for the acceleration probability density (of the moving source) whose mean value is the current acceleration. Based on this model, we utilize a zoom factor achieved by computing the norm of the tracking innovation to adjust the dynamic state model adaptively. Both numerical simulation and the outdoor experimental results show that the proposed framework in this paper can effectively track a maneuvering speech source.