Visual speaker identification with spatiotemporal directional features

Abstract

In this paper, a novel local spatiotemporal directional descriptor is proposed for speaker identification by analyzing mouth movements. For this new descriptor, the directional local binary pattern features in three orthogonal planes are coded. In addition, besides sign features, magnitude information encoded as weight for the bins with the same sign value is developed to improve the discriminative ability. Moreover, decorrelation is exploited to remove the redundancy of features. Experimental results on the challenging XM2VTS database show the effectiveness of the proposed representation for this problem. © 2013 Springer-Verlag.