Abstract: For the speaker classification issue with missing priori information, it could be used to extract the spatial acoustic features based on multiple distance microphone to classify speakers. Its dimension needed to be reduced to solve the problem of computational cost caused by the rapid growth of space acoustic features dimension with the increasing microphone's number. A novel optimized discriminant locality preserving projections speaker classification method was proposed, which could preserve manifold structure of spatial acoustic feature and decrease computing cost. Experiments were validated with the speech conference data set collected by the multi-distance microphones. In most of the data set the diarization error rate (DER) score of this method was lower than traditional methods. Results show that this method has the improved speaker classification performance than the traditional methods.