Speaker identification based on fuzzy-neural networks

In this paper, F ratio formula, which is used to measure the effectiveness of the parameters for representing speaker individual features, is employed to compare the effectiveness of LSP frequencies with that of CEP parameters. Some specific properties have been drawn from the so two kinds of parameters. Based on this work, LSP frequencies are determined as individual features. Then a hierarchical model for describing individual features of each speaker is proposed. According to the capability of interpolation with functional——link networks and the fuzzy statistical method of establishing membership functions, the membership functions of the fuzzy states are implemellted with such networks which are the basic units of the hierarchical model. During the identification procedure, the maximum degree of membership fUnction to speaker's model is used as the decision criterion. In small fixed vocabulary, ten Chinese digital voices (0-9), text-independent speaker identification test are carried out among 42 speakers. When the length of testing utterances is randomly concatenated with 5 digital voices, the correct identification rate is 99.76 percent.