Combining dynamic texture and structural features for speaker identification.

Visual information from captured video is important for speaker identification under noisy conditions that have background noise or cross talk among speakers. In this paper, we propose local spatiotemporal descriptors to represent and recognize speakers based solely on visual features. Spatiotemporal dynamic texture features of local binary patterns extracted from localized mouth regions are used for describing motion information in utterances, which can capture the spatial and temporal transition characteristics. Structural edge map features are extracted from the image frames for representing appearance characteristics. Combination of dynamic texture and structural features takes both motion and appearance together into account, providing the description ability for spatiotemporal development in speech. In our experiments on BANCA and XM2VTS databases the proposed method obtained promising recognition results comparing to the other features.