Improving performance of text-independent speaker identification by utilizing contextual principal curves filtering

In this paper, a novel filtering method in feature extraction of speech is proposed for text-independent speaker identification, called Contextual Principal Curves Filtering (CPCF). The CPCF provides a good nonlinear summary of a sequence of cepstral vectors on the time context and, the most important, keeps their intrinsic trajectory characteristics, so the CPCF algorithm do improve the cepstral coefficients to represent speech feature more precisely. We apply this CPCF algorithm into two protocols in the framework of close-set textindependent speaker identification, where the experimental data are collected from a subset of 863 speech database of China National High Technology Project. The results show a steady relative error rate reduction of the identification for more than 20% compared with the use of the conventional Mel-frequency cepstral coefficients under both of the two protocols.