As an improvement to cepstral coefficients, an approach, called variance-based filtering model (VTM) of speech, is presented in this paper, in which the information of cepstral vectors is decomposed into intrinsic and noise components according to their variances. And then, contextual principal curves filtering (CPCF) is introduced as an algorithm of the VFM model, and this curve provides good nonlinear summary of the cepstral vectors and keeps their intrinsic trajectory characteristics. We finally apply this CPCF algorithm in the framework of speaker identification, using a subset of 863 speech database of China National High Technology Project. The results show a relative improvement of roughly 27% compared to the use of the classical cepstral coefficients augmented by their Delta coefficients