Angle of Models Distance as Test Algorithm in Speaker Verification
Wu Guo, Pan Lei, Ren-Hua Wang, Li-Rong Dai · 2007
In this paper we introduce the angle of models Mahalanobis distance as the test algorithm in the text- independent speaker verification system. In test process, the test speech is adapted to a new model instead of calculating the log-likelihood ratio scores, then the Mahalanobis distances among the UBM model, the speaker model and the test speech model are calculated. These three models distances form a triangle. The angle of the triangle can be treated as the test scores. When we employ the proposed algorithm in GMM_UBM system, the recognition rate can be almost identical with the traditional log-likelihood ratio scores, while the computation load can be dramatically cut down. That is very helpful to the real-time application of the speaker recognition or verification. Further more, when we fuse angle scores with the log-likelihood ratio scores, the EER of the speaker verification can be reduced by 12-15% in 2002 NIST evaluation corpus.