The performance comparison of fitting feature with segment model in speaker identification

In general, cepstra are used to manifest the shape of the vocal tract. But it ignores the various stochastic errors of the vocal tract, which can be observed from different speech signals is obtained from the same word pronounced repeatedly even if ignoring the noise and the speaking rate. Now we make use of a method to eliminate the various stochastic errors to some extent, and the feature vectors can represent the essential characteristics of the vocal tract more effectively. In this paper, we analyze the speech-production process and utilize a segment model to implement the elimination of the various stochastic errors. The experiments show: besides the elimination of the stochastic errors, the proposed method also can remove the noise in the environment.