On the use of autoregressive modeling for localization of speech

Abstract

The localization of speech is essential for improving the quality of hands-free pick-up as well as for applications such as automatic camera steering. This paper proposes a source localization method tailored to the distinct nature of speech that is based on the linearly constrained minimum variance (LCMV) beamforming method. The LCMV steered beam temporally focuses the array onto the desired signal. By modeling the desired signal as an autoregressive (AR) process and embedding the AR coefficients in the linear constraints, the localization accuracy is significantly improved as compared to existing techniques.