Speaker classification is an important component for audio indexing technology for many applications such as multimedia conferencing. The primary input device of NIST speaker classification evaluation is Multiple Distant Microphones (MDM). MDM is composed of multiple microphones and has the merit of low price and easy-to-use. The spatial time-delay vector of MDM can be extracted as the speaker's discriminant feature. However the feature dimension will be expanded quickly with the increasing number of sensors. Locality Preserving Projections (LPP) and Discriminant locality preserving projection (DLPP) are the principal manifold dimension-reduction algorithms being proposed recently. In this paper, we proposed a novel method to overcome the drawbacks of traditional manifold algorithms such as the lack of class information or spatial identification information. Some basic concepts of spatial time-delay feature and merging feature for MDM speaker classification are first introduced. A review of known DLPP algorithm followed by Fisher criterion is given. Then the Multi-component Discriminant Locality Preserving Projections (MDLPP) method for speaker classification with MDM is described. Comparative experiment results on real meeting data showed the effectiveness of the proposed method.
展开▼
