New speaker recognition feature using correlation dimension

Abstract

The authors propose a correlation dimension as a new feature vector of a speaker verification system among the persons whose voices are much the same as each other. Since the speech production mechanism is nonlinear, the ambiguity of speech utterance can be regarded as a chaotic property. For speakers whose voices appear to be nearly identical, using their vowel sound /i/ which is more chaotic than /a/, /e/, /o/, or /u/ (in Korean), the authors carry out speaker verification experiments and analyze the efficiency of speaker verification with CHMM (continuous hidden Markov model), formant and correlation dimension.