Our goal was to build a text-independent speaker recognition system that could be used under any conditions without any additional adaptation process. Unknown mismatched microphones and noise conditions can severely degrade the performance of speaker recognition systems. This paper shows that principal component analysis (PCA) can increase performance under these conditions without reducing dimension. We also propose a PCA process that augments class discriminative information sent to original feature vectors before PCA transformation and selects the best direction between each pair of highly confusable speakers. In tests, the proposed method reduced errors in recognition by 32%.
展开▼
