Application of KPCA and PNN for Robust Speaker Identification

This paper presents a robust speaker identification approach basing on kernel principle component analysis (KPCA) and probabilistic neural network (PNN). KPCA is exploited to reduce the dimension of input vector and to denoise speech signal by extracting the nonlinear principle components of the feature vector. The extracted principle components are utilized as the input feature vector of the classifier and a probabilistic neural network (PNN) is designed as the classifier of identification system. We have tested our system on KING corpus and the experimental result shows that our system outperforms PNN and GMM approach in terms of robustness and training time.