Eigen-MLLR Coefficients as New Feature Parameters for Speaker Identification

Eigen-MLLR coefficients are proposed as new feature pa- rameters for speaker-identiffication in this paper. By performing principle component analysis on MLLR parameters among training speakers, the eigen-MLLR coefficients (EMCs) are derived as the coefficients for the eigenvectors. The discriminating function of the new EMC features based on the Fisher criterion is found to be ten times larger than that of mel-frequency cepstral coefficient (MFCC) features, for distinguishing speakers. The speaker-identification accuracy using the EMC features are shown to be significantly better than that using MFCC features, especially when the quantity of enrollment data is limited. It is also shown that properly combining MFCC and EMC features can achieve a significant error rate re- duction on the order of 50%-60% as compared to using MFCC features alone.